chunk_uid
stringlengths
40
40
chunk_type
stringclasses
2 values
chunk_index
int64
0
6.71k
total_chunks
int64
1
6.71k
section_title
stringlengths
1
157
embed_text
stringlengths
1
83.3k
spans
dict
paper_doi
stringlengths
0
63
paper_id_arxiv
stringlengths
9
16
title
stringlengths
7
245
authors
listlengths
1
768
categories
listlengths
1
7
year
int64
2k
2.02k
language
stringclasses
2 values
discipline
stringclasses
8 values
dense_vector
listlengths
1.02k
1.02k
bddf07f7d59fcb0aa9c6e3173611a43f94b03d1f
subsection
7
37
Wrapping up
In Narcissus, users specify formats using a library of combinators, and use tactics to automatically derive correct-by-construction encoder and decoder functions from these specifications. Formats may be underspecified, in that a particular source value may be serialized in different ways, but decoders are guaranteed t...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/978-3-642-00590-9_12", "end": 2201, "openalex_id": "https://openalex.org/W1509784638", "raw": "Aditi Barthwal and Michael Norrish. 2009. Verified, Executable Parsing. In Programming Languages and Systems, Giuseppe Castagna (Ed.). Sp...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.0033734957687556744, -0.008825629018247128, -0.03234283998608589, 0.03777399659156799, 0.007444955874234438, -0.02637772262096405, 0.02311292663216591, -0.012685411609709263, 0.005713393446058035, 0.06578411906957626, -0.04662251099944115, -0.012281126342713833, 0.026057345792651176, -0....
90245bd9840be423c2a28b9f0c81e639f13b9bbe
subsection
8
37
Narcissus, Formally
We begin our ground-up explanation of Narcissus with the definition of the formats that capture relationships between structured source values and their serialized representations. The signature of a format from source type S to target type |T| is defined by a type alias: FormatM S T BIGSIGMA := Set of (S * BIGSIGMA *...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.00345655158162117, -0.02228064276278019, -0.05295467749238014, 0.027164069935679436, 0.020830875262618065, -0.058112796396017075, 0.018404420465230942, -0.00827130675315857, 0.006314120255410671, 0.06293518096208572, 0.02090717852115631, 0.0010997251374647021, 0.024676574394106865, -0.0...
f31a5b0c4618a3f5a0332b224f1ed9f54d2ca9c2
subsection
9
37
Narcissus, Formally
The three operators have straightforward interpretations as sets : e ELEMENT return v === e = v e ELEMENT x | P x === P e e ELEMENT x <- y; k x === exists e'. e' ELEMENT y / e ELEMENT k e'As an example, we can specify the set of all possible locations of a period in a string s as: s1 <- s1 : String | exists s2. s = s...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 414, "openalex_id": "", "raw": "Benjamin Delaware, Clément Pit-Claudel, Jason Gross, and Adam Chlipala. 2015. Fiat: Deductive Synthesis of Abstract Data Types in a Proof Assistant. In Proc. POPL.", "source_ref_id": "ab5ef2f2...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.020555714145302773, -0.02788069099187851, -0.06232333928346634, 0.03143635764718056, 0.013192587532103062, -0.04846692457795143, 0.055639300495386124, -0.016939006745815277, 0.021272951737046242, 0.0750199630856514, -0.011338452808558941, -0.0017358667682856321, 0.03534300997853279, -0.0...
5ad328f74e41fbb6fdc41d380c92167270fa231c
subsection
10
37
Specifying Encoders and Decoders
These relational formats are not particularly useful by themselves— even checking whether a format permits specific source and target values may be undecidable. Instead we use them to specify the correctness of both encoders and decoders. So far, we have seen examples of relational formats that permit one or many targe...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.028401220217347145, 0.012102185748517513, -0.05341443791985512, 0.020358530804514885, 0.009828256443142891, -0.03516196087002754, 0.028080733492970467, -0.037359584122896194, -0.0030274540185928345, 0.0372374951839447, -0.0003338402311783284, -0.004677578806877136, 0.030308878049254417, ...
e51d50b419682b9eccb274b95b722f1fd8aeb901
subsection
11
37
Specifying Encoders and Decoders
For now, we require correct decoders to flag strictly all malformed target values by signaling errors when applied to target values not included in the relation: \forall s \, t, \; \mathsf {decode}\; t = \mathsf {Some}\; s \rightarrow (s, t) \in \mathsf {format}. As we shall see later, our formulation will also support...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 1579, "openalex_id": "https://openalex.org/W2111427271", "raw": "P. Mockapetris. 1987. Domain names - implementation and specification. RFC 1035.", "source_ref_id": "fa897e93f3ceaed0b4d52159bfbafc8bc39d9758", "start": ...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.011657718569040298, 0.005279542412608862, -0.049560561776161194, 0.022979743778705597, 0.014877323992550373, 0.0034294137731194496, 0.015281681902706623, -0.039764415472745895, 0.0007276537362486124, 0.050872817635536194, -0.0009679797221906483, -0.015296940691769123, 0.037048351019620895,...
8db350c4e3b37c8ba0cc323728b87a1df11ee3d2
subsection
12
37
Specifying Encoders and Decoders
Hence our full notion of decoder correctness accounts for both state and erroneous target values.[Decoder Correctness] A correct decoder for a format, |format : FormatM S T BIGSIGMAE|, and relation on states, |EQUIV : Set of (BIGSIGMAE * BIGSIGMAD)|, is a function, |decode : DecodeM S T BIGSIGMAD|, that, when applied t...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.005786298308521509, -0.01362469233572483, -0.06426338851451874, -0.013121204450726509, 0.014654554426670074, -0.0019338518613949418, -0.009505243971943855, -0.005824441555887461, 0.0007695357198826969, 0.021512672305107117, -0.00509972358122468, -0.009078042581677437, 0.0105579923838377, ...
2d19609d7be99acb89dab6f906be96cc18ae509c
subsection
13
37
Specifying Encoders and Decoders
\; \mathsf {decode}\; t\;\sigma _D = \mathsf {Some}(s, \sigma _D^{\prime })[Encode Inverts Decode] Given a correct decoder \mathsf {format} \hspace{-2.84544pt}\begin{}[baseline=-0.63ex, ->, thick]\node (A) { };\node (B) [right of=A, node distance=3em] { };\node (C) [right of=A, node distance=1.5em] {\approx };[every n...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.0037437633145600557, 0.016401613131165504, -0.026105264201760292, -0.012587283737957478, 0.013060260564088821, -0.0019605648703873158, -0.0068162051029503345, 0.014715678989887238, 0.02853117696940899, 0.01801888830959797, -0.0013989049475640059, 0.022962257266044617, -0.019437817856669426...
77f16bb4ec124861e6e3dda81764eecd7c1746e7
subsection
14
37
Deriving Encoders and Decoders
Equipped with precise notions of correctness, we can now define how we derive provably correct encoders and decoders from a format. These functions will be byte-aligned in a subsequent derivation step presented in sec:ByteAlignment. We begin with encoders, since they often have similar structure to their corresponding ...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.03525542840361595, 0.00918777845799923, -0.0403224416077137, 0.03537752479314804, -0.0016692911740392447, -0.034736517816782, -0.0004654937656596303, -0.005395148880779743, -0.024922993034124374, 0.03684268519282341, -0.0291200689971447, -0.018345031887292862, -0.002270235912874341, 0.0...
61a59131de7e45504890c41a5e9b6ed0c41e981e
subsection
15
37
Decoders
Before defining similar correctness rules for decoder combinators, we pause to consider how they are used to build a top-level decoder. In particular, consider what the decoder combinators used to build a reusable decoder for |++| should look like:\mathsf {format_1 +\hspace{-3.0pt}+\, format_2}~ \hspace{-2.84544pt}\beg...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.016209140419960022, 0.017628584057092667, -0.0446590818464756, 0.02608419954776764, 0.024664755910634995, -0.02387108840048313, 0.006143294740468264, -0.009089023806154728, -0.0143623361364007, 0.016789129003882408, -0.028114158660173416, 0.0008098846301436424, -0.004658983089029789, -0....
35b3abeb28638f61b9a0bef4d356de63b21c4eaa
subsection
16
37
Decoders
In order to account for the first two concerns, our adaptation of the soundness (left-inverse) criterion for decoder combinators is parameterized over a binary relation on source and projected values representing a view, as well as an additional format capturing the conformance checking performed by the decoder.[Decode...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.012461507692933083, -0.012583605013787746, -0.0802176222205162, 0.007367780897766352, 0.0006271770689636469, -0.02408359944820404, 0.01081320084631443, -0.01984073594212532, -0.005002154503017664, 0.0434664711356163, -0.012850691564381123, -0.02901325933635235, 0.01984073594212532, -0.0...
3af2fb14a77bc6ed3232de4f924f4d5d25f023ab
subsection
17
37
Decoders
Absent a complete view of the original source value, a combinator will be unable to ensure adherence to the original format, but it can ensure that any computed value agrees with the provided view format relation and is a consistent view of any source values in the original format with the same encoding: [Decoder Combi...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.006503456272184849, -0.02168073132634163, -0.04577212780714035, 0.007026021368801594, 0.010550475679337978, -0.025189926847815514, -0.0035702260211110115, -0.008902679197490215, -0.0004558140935841948, 0.043300431221723557, -0.0014008177677169442, -0.029080558568239212, 0.00756384385749697...
dd05752fcacaa53a3a6fd68e25951ef718783d79
subsection
18
37
Decoders
This predicate is key to our approach to the modular verification of decoder combinators: each of these proofs use this predicate to thread information about previously decoded data through a proof of correctness for a composite decoder. Such information is necessary for a decoder combinator whose correctness depends o...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.013753163628280163, -0.012280972674489021, -0.05470753833651543, 0.015972891822457314, 0.009550172835588455, -0.03429518640041351, -0.012990370392799377, -0.006590534467250109, -0.010610454715788364, 0.019786858931183815, -0.03859734162688255, -0.03447825834155083, 0.002030937233939767, ...
550c6027385cdeb774a922795e6252991250f4d7
subsection
19
37
Decoders
While not particularly helpful during derivations, this rule is useful for proving other derivation rules. DecViewDone generalizes DecDone to arbitrary views of a source value. The second premise corresponds to the decision procedure from DecDone — an empty conformance format is one consequence of the source value pro...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.009255764074623585, -0.006398152094334364, -0.009499939158558846, 0.009553352370858192, 0.002867150353267789, -0.015795079991221428, 0.02328820526599884, -0.018297875300049782, 0.005440527573227882, 0.04459248483181, -0.035191744565963745, -0.04141820967197418, 0.011590689420700073, 0.0...
fd849e4341435aaf91bd21d26c9a2f1bd573b271
subsection
20
37
Improving Performance of Encoders and Decoders
The encoders and decoders derived via our combinator rules utilize the same bitstring abstract data type as format specifications, employing the bitstring's |snoc| and |unfold| operations to enqueue and dequeue individual bits. Operating at the bit-level imposes a large performance hit on these functions, since impleme...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.0146162249147892, -0.017301460728049278, -0.06761910766363144, 0.002799662994220853, 0.0130600081756711, 0.026242073625326157, 0.021146228536963463, -0.013487204909324646, -0.0022999956272542477, 0.024609573185443878, -0.03744072467088699, -0.02839331328868866, 0.019269615411758423, -0.0...
bff49b439a6815a054159a67f3528c158210eff8
subsection
21
37
Improving Performance of Encoders and Decoders
We define the twin equivalences used to justify the correctness of byte-optimized functions as follows: [Correctness of Byte-Aligned Encoders] A byte-aligned encoder |encodebytes| and bit-aligned encoder |encodebits| are equivalent, \mathsf {encode\_bits \simeq encode\_bytes}, iff:|encodebytes| encodes the same bit seq...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.039920952171087265, 0.009675032459199429, -0.05103045329451561, 0.014916947111487389, 0.008850975893437862, 0.006973958574235439, 0.018159762024879456, 0.001134985126554966, 0.007881946861743927, 0.0317109115421772, -0.06256724148988724, -0.02980337291955948, -0.012246393598616123, 0.009...
57a72e711d7fddacf5e5211a87b8e197d88996ad
subsection
22
37
Automating Derivations
As illustrated in sec:narcissusTour, Narcissus provides a set of tactics to help automate the derivations described above. The tactic derive_encoder_decoder_pair presented in that tour is actually implemented via a pair of proof-automation tactics, |DeriveEncoder| and |DeriveDecoder|, that derive encoders and decoders,...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.008034946396946907, -0.024402430281043053, -0.048194415867328644, 0.019808851182460785, 0.0045477948151528835, 0.0014393083984032273, 0.03320806100964546, -0.035771917551755905, 0.0341847687959671, 0.06501210480928421, -0.04266991466283798, 0.01140763983130455, 0.0015480434522032738, 0....
6b9f53d0e5eebd87b6f85bff151b52e4abffc7e2
subsection
23
37
Automating Derivations
The subformats represent the \begin{}[baseline=-0.63ex, ->, thick]\node (A) { };\node (B) [right of=A, node distance=3em] { };\node (C) [right of=A, node distance=1.5em] {\approx }; \end{}[every node/.style={font=\sffamily \small }] (A) edge[bend left=20] node [right] {} (B) (B) edge[bend left=20] node [left] {} (A); p...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.04968041554093361, -0.005447146017104387, -0.04305839166045189, -0.01745527982711792, -0.007556580007076263, 0.019774513319134712, 0.025954049080610275, -0.000019549455828382634, -0.020613709464669228, 0.020323805510997772, -0.0272204726934433, -0.002715944079682231, -0.009994063526391983...
29093e52ed89507ad2971017a75609d085223a2a
subsection
24
37
Automating Derivations
Automatically finding an instance of \mathsf {s} is particularly worrisome, as it is well-known that Ltac, Coq's proof-automation language, does not provide good support for introspecting into definitions of inductive types, and we would like to use Ltac to construct records of fairly arbitrary types, without relying o...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.03812701255083084, 0.006410466507077217, -0.019719816744327545, 0.013454348780214787, 0.01110384427011013, -0.004575089085847139, 0.0533289760351181, -0.009653857909142971, 0.02246715873479843, 0.06416571885347366, -0.007780322339385748, -0.011531208641827106, -0.004567457363009453, 0.0...
f0194821c12287399674077ae7b8511c29c8215a
subsection
25
37
Extending the Framework
As outlined in sec:narcissusTour, an extension to Narcissus consists of four pieces: a format, encoder and decoder combinators, derivation rules, and automation for incorporating these rules into DeriveEncoder and DeriveDecoder. As a concrete example, consider the format of the Internet Protocol (IP) checksum used in t...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.008482773788273335, -0.018948929384350777, -0.029613422229886055, 0.025722941383719444, 0.000951165275182575, -0.022213881835341454, 0.01088571734726429, -0.021649379283189774, -0.015653463080525398, 0.05364286154508591, -0.040308430790901184, 0.003282116260379553, -0.0017249704105779529,...
b3fbf4c29237f28297ab5249f9e955869f61ed7f
subsection
26
37
Evaluation
To evaluate the expressiveness and real-world applicability of Narcissus, we wrote specifications and derived implementations of encoders and decoders for five of the most commonly used packet formats of the Internet protocol suite: Ethernet, ARP, IPv4, TCP, UDP. These formats were chosen to cover the full TCP/IP stack...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.003382350318133831, 0.009744525887072086, -0.026295723393559456, 0.019443267956376076, 0.01779501512646675, -0.041725192219018936, 0.015215806663036346, -0.021412013098597527, 0.026662001386284828, 0.04999697580933571, -0.02676883153617382, 0.009401140734553337, 0.02031317912042141, -0....
d24cfb866bf7961920ae986e260f39f96eb61732
subsection
27
37
Benchmarking
fig:benchmarks shows single-packet encoding and decoding times, estimated by linearly regressing over the time needed to run batches of n packet serializations or deserializations for increasingly large values of n (complete experimental data, including 95% confidence intervals, are provided as supplementary material; ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 376, "openalex_id": "", "raw": "Christopher S. Hardin and Roshan P. James. 2013. Core_bench: micro-benchmarking for OCaml. (2013). https://github.com/janestreet/core_bench", "source_ref_id": "5ed4250ef8dcef0a1bca024588311c5c...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.018667027354240417, -0.004590440541505814, -0.029946614056825638, -0.007402705028653145, 0.007898762822151184, -0.03409823402762413, 0.0033006907906383276, 0.003731879172846675, -0.032022424042224884, 0.027199218049645424, -0.012851706705987453, -0.021399160847067833, -0.02402444928884506...
3123ceeb79cb33fdafc76f9bdeb3607d7ca36599
subsection
28
37
Mirage OS Integration
MirageOS  is a “library operating system that constructs unikernels for secure, high-performance network applications”: a collection of OCaml libraries that can be assembled into a standalone kernel running on top of the Xen hypervisor. Security is a core feature of MirageOS, making it a natural target to demonstrate i...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 236, "openalex_id": "", "raw": "Anil Madhavapeddy, Richard Mortier, Charalampos Rotsos, David Scott, Balraj Singh, Thomas Gazagnaire, Steven Smith, Steven Hand, and Jon Crowcroft. 2013. Unikernels: Library Operating Systems for th...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.004715463146567345, -0.0023939749225974083, -0.037357453256845474, -0.017381593585014343, 0.010964595712721348, -0.028094392269849777, 0.014161649160087109, 0.003057802328839898, -0.0068213981576263905, 0.01858716458082199, -0.03171110898256302, -0.039615992456674576, -0.02527122013270855...
3bb9636fda037b3b05107882f6f41c36e1ade760
subsection
29
37
Setup
After extracting the individual encoders and decoders to OCaml, we reprogrammed the TCP, UDP, IPv4, ARPv4, and Ethernet modules of the mirage-tcpip library to use our code optionally, and we recompiled everything. This whole process went smoothly: Mirage's test suite did not reveal issues with our proofs, though we did...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.027047403156757355, -0.002271112287417054, -0.003836673451587558, -0.002234881278127432, 0.019648650661110878, -0.036703921854496, -0.0015550731914117932, -0.003672680351883173, -0.004008294083178043, 0.035086873918771744, -0.02018257975578308, 0.005846541374921799, -0.02038089744746685, ...
ac0dff06e3464d8d33ec79c29badc6ff0e9118dc
subsection
30
37
Parsers for Context-Free Languages
There is a long tradition of generating parsers for context-free languages from declarative Backus-Naur-form specifications , automatically. Such generators may themselves have errors in them, so in order to reduce the trusted code base of formally verified compilers, there have been a number of efforts in verifying st...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 141, "openalex_id": "", "raw": "Stephen C. Johnson. 1979. Yacc: Yet Another Compiler-Compiler. Technical Report.", "source_ref_id": "5169855be148f4cd6c6c5daa720ba3003b0ee701", "start": 0 }, { "arxiv_id": ...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.019225358963012695, -0.031523484736680984, -0.03756574168801308, 0.03668076545000076, 0.008681927807629108, 0.004108276218175888, 0.03143193572759628, -0.0024260571226477623, 0.0073201316408813, 0.03001292236149311, -0.017440147697925568, -0.021819256246089935, -0.002223885850980878, -0...
6fc721cc19d02f47278e2b252edd0df9c34d49b9
subsection
31
37
Verification of Parsers for Network Protocol Formats
A wide range of tools have been used to verify generated parsers for binary network protocol formats , , , , , including the SAW symbolic-analysis engine,  , the Frama-C analyzer , F* , Agda , and Coq. The correctness properties of each project differ from Narcissus's: focus on memory safety. While and prove that a p...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/978-3-319-96142-2_25", "end": 201, "openalex_id": "https://openalex.org/W2883286797", "raw": "Nathan Collins, Mark Tullsen, Aaron Tomb, and Lee Pike. 2017. Formal Verification of a Vehicle-to-Vehicle (V2V) Messaging System. In Embed...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.028244324028491974, -0.003772776573896408, 0.010337941348552704, 0.008560273796319962, -0.014076385647058487, -0.04696706309914589, 0.04852347448468208, -0.013908537104725838, -0.004940086510032415, 0.06463693082332611, -0.016449153423309326, -0.00003135238148388453, 0.06024235114455223, ...
fdc540cf16e8e406e9675a9ed9502197e75cb2db
subsection
32
37
Deductive Synthesis
The idea of deriving correct-by-construction implementations from specifications using deductive rules has existed for at least half a century , . Kestrel's Specware  system was an seminal realization of this idea, and has been used to implement correct-by-construction SAT solvers , garbage collectors , and network pro...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/bf01933419", "end": 146, "openalex_id": "https://openalex.org/W2078250023", "raw": "Edsger W. Dijkstra. 1967. A constructive approach to the problem of program correctness. (Aug. 1967). http://www.cs.utexas.edu/users/EWD/ewd02xx/EWD...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.028271647170186043, -0.015974320471286774, -0.014921571128070354, 0.05166098102927208, 0.0005416316562332213, -0.03031611628830433, 0.030422916635870934, -0.023831794038414955, -0.021665267646312714, 0.05895393714308739, -0.014173966832458973, -0.020795604214072227, 0.012739786878228188, ...
5c8be60cc7d7aa4e32f94ddd0ca15b8692e9085b
subsection
33
37
Parser-Combinator Libraries
There is a long history in the functional-programming community of using combinators  to eliminate the burden of writing parsers by hand, but less attention has been paid to the question of how to generate both encoders and decoders. Kennedy  presents a library of combinators that package serializers and deserializers ...
{ "cite_spans": [ { "arxiv_id": "", "doi": "", "end": 233, "openalex_id": "https://openalex.org/W2123835026", "raw": "Daan Leijen and Erik Meijer. 2001. Parsec: Direct style monadic parser combinators for the real world. (2001).", "source_ref_id": "5f52b998f39eac493ae3b43695239...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.03165166452527046, -0.009042242541909218, -0.03281151130795479, 0.01974795199930668, 0.016848329454660416, -0.02296805940568447, 0.02624920941889286, -0.0326283797621727, -0.007619138341397047, 0.012682030908763409, -0.01297962386161089, 0.002546707633882761, -0.009996065869927406, -0.0...
005cc7fc1a4123bebccff07601a48374a603ed68
subsection
34
37
Bidirectional / Invertible Programming Languages
Mu et. al present a functional language in which only injective functions can be defined, allowing users to invert every program automatically . The authors give a relational semantics to this language, although every program in the language is a function. The authors show how to embed noninjective programs in their la...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.1007/978-3-540-27764-4_16", "end": 144, "openalex_id": "https://openalex.org/W2115810994", "raw": "Shin-Cheng Mu, Zhenjiang Hu, and Masato Takeichi. 2004. An Injective Language for Reversible Computation. In Mathematics of Program Constr...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.013441607356071472, -0.011191167868673801, -0.0016678044339641929, 0.013777266256511211, 0.012694003991782665, -0.07305154949426651, 0.014585898257791996, -0.0017192974919453263, 0.013601807877421379, 0.03893641382455826, -0.03789892420172691, -0.02372497133910656, 0.04479518532752991, ...
257c463808626e836411071e854b17dd99d15cd4
subsection
35
37
Extensible Format-Description Languages
Interface generators like XDR , ASN.1 , Apache Avro , and Protocol Buffers  generate encoders and decoders from user-defined data schemes. The underlying data format for these frameworks can be context-sensitive, but this format is defined by the system, however, preventing data exchange between programs using differen...
{ "cite_spans": [ { "arxiv_id": "", "doi": "10.17487/rfc4506", "end": 36, "openalex_id": "https://openalex.org/W1700387131", "raw": "Raj Srinivasan. 1995. XDR: External data representation standard. Technical Report.", "source_ref_id": "18c909cc2a81663eb7d85c0ed4d0d30448b71b39"...
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ -0.03846188634634018, -0.011622510850429535, -0.007921317592263222, 0.0325399786233902, -0.014842930249869823, -0.03522620350122452, 0.02973164990544319, 0.012492481619119644, -0.014560571871697903, 0.036111436784267426, -0.008493666537106037, -0.0205740574747324, 0.001976513536646962, -0....
1ccd3d047955bcb387d2514b85ff91b83c7a9312
subsection
36
37
Conclusion
We have presented Narcissus, a framework for specifying and deriving correct-by-construction encoders and decoders for non-context-free formats in the style of parser-combinator libraries. This framework provides fine-grained control over the shape of encoded data, is extensible with user-defined formats and implementa...
{ "cite_spans": [] }
1803.04870
Narcissus: Deriving Correct-By-Construction Decoders and Encoders from Binary Formats
[ "Benjamin Delaware", "Sorawit Suriyakarn", "Clément Pit--Claudel", "Qianchuan Ye", "Adam Chlipala" ]
[ "cs.PL" ]
2,018
en
Computer Science
[ 0.0058520580641925335, -0.000402472069254145, -0.0250104907900095, 0.012772288173437119, 0.007481020875275135, -0.024812115356326103, 0.041078854352235794, -0.01988326385617256, 0.01683134399354458, 0.06604356318712234, -0.01182619296014309, -0.029664671048521996, 0.025987105444073677, -0....
828e02b8f4fa31cdad610ae1da33a6c132b44a61
abstract
0
160
Abstract
We examine a class of deep learning models with a tractable method to compute information-theoretic quantities. Our contributions are three-fold: (i) We show how entropies and mutual informations can be derived from heuristic statistical physics methods, under the assumption that weight matrices are independent and ort...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.00017650832887738943, -0.034446414560079575, 0.008771784603595734, 0.017833419144153595, -0.008298871107399464, -0.08616180717945099, 0.03353109955787659, 0.033195484429597855, -0.008565838448703289, 0.023645680397748947, -0.033927738666534424, -0.017528314143419266, -0.013966206461191177...
2889f71ed66cd43b6f03dfa2ba078fac54e7abf1
subsection
1
160
Multi-layer model and main theoretical results
A stochastic multi-layer model— We consider a model of multi-layer stochastic feed-forward neural network where each element x_i of the input layer {x} \in \mathbb {R}^{n_0} is distributed independently as P_0 (x_i), while hidden units t_{\ell , i} at each successive layer {t}_\ell \in \mathbb {R}^{n_{\ell }} (vectors ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.023548929020762444, -0.002045078668743372, -0.04004691541194916, -0.006501519121229649, -0.008111637085676193, 0.019046703353524208, 0.014170564711093903, 0.06715183705091476, 0.011827880516648293, 0.004925739951431751, -0.02730332687497139, 0.014162933453917503, -0.03617042303085327, 0...
8c66becef99902cdf6646a8d890f9f6feb47fe35
subsection
2
160
Multi-layer model and main theoretical results
Model (REF ) thus describes a Markov chain which we denote by \rightarrow 1 \rightarrow 2 \rightarrow \dots \rightarrow L, with {\ell }=\varphi _{\ell }(W_{\ell } {\ell -1}, { {\xi }}_\ell ), {\xi }_\ell =\lbrace {\xi }_{\ell ,i}\rbrace _{i=1}^{n_\ell }, and the activation function \varphi _\ell applied componentwise.R...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.008886218070983887, -0.0033771442249417305, 0.0034648622386157513, -0.019206436350941658, -0.0033485405147075653, 0.007478915620595217, 0.041463933885097504, 0.04631512239575386, -0.0008709828834980726, 0.0040579126216471195, -0.056658223271369934, 0.04543031379580498, -0.0027078175917267...
fc998e9e159a595b5586cf68f06f993bb5bbc7c3
subsection
3
160
Multi-layer model and main theoretical results
Then for any \ell \in \lbrace 1,\ldots ,L\rbrace the normalized entropy of \ell is given by the minimum among all stationary points of the replica potential:\lim _{n_0 \rightarrow \infty } \frac{1}{n_0} H({\ell }) = \min _{{A}, {V}, {\tilde{A}}, {\tilde{V}}} \phi _\ell ({A}, {V}, {\tilde{A}}, {\tilde{V}}),which depends...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05917533114552498, 0.010628060437738895, 0.0038052727468311787, 0.009727764874696732, -0.015884868800640106, -0.023239821195602417, 0.0102465795353055, 0.013725685887038708, 0.017761753872036934, 0.005985436495393515, -0.02797018550336361, 0.01150546595454216, 0.01818901300430298, 0.020...
18d1e0a01c18b359bd4756088f6b77de083a02cb
subsection
4
160
Multi-layer model and main theoretical results
In the computation of the conditional entropies in (REF ), the scalar t_k-variables are generated from P(t_0)=P_0(t_0) andP(t_{k} | \xi _k; A, V, \rho ) &= \mathbb {E}_{\tilde{\xi }, \tilde{z}} \, P_k (t_k + \tilde{\xi } / \sqrt{A} | \sqrt{\rho - V} \xi _k + \sqrt{V} \tilde{z}), \quad k=1,\dots ,\ell -1,\\ P(t_{\ell } ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0487801693379879, 0.008119853213429451, -0.0068644811399281025, -0.002064304891973734, 0.006398963276296854, -0.002930473303422332, 0.04038558527827263, 0.01588865928351879, 0.014881310053169727, 0.026053739711642265, -0.03919507935643196, 0.02439008466899395, -0.03837088495492935, 0.04...
01d0bea128188b1152bf4daf2437a3b0f05da646
subsection
5
160
Multi-layer model and main theoretical results
As ours, it exhibits layer-wise additivity, and the two formulas are conjectured to be equivalent.Rigorous statement— We recall the assumptions under which the replica formula of Claim is conjectured to be exact: (i) weight matrices are drawn from an ensemble of random orthogonally-invariant matrices, (ii) matrices at...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.022918781265616417, 0.016906797885894775, -0.0005974790547043085, 0.005634327884763479, 0.00683596171438694, -0.015434319153428078, 0.008438140153884888, 0.030853379517793655, -0.0022602160461246967, 0.018814153969287872, -0.04003920406103134, 0.013946582563221455, -0.03973402827978134, ...
cad5fc5ed633104ca12fc3e96185702bcd176289
subsection
6
160
Tractable models for deep learning
The multi-layer model presented above can be leveraged to simulate two prototypical settings of deep supervised learning on synthetic datasets amenable to the replica tractable computation of entropies and mutual informations. [Figure: Two models of synthetic data]The first scenario is the so-called teacher-student (se...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03783775866031647, -0.009703554213047028, -0.018323849886655807, -0.003861968871206045, 0.02294677123427391, -0.026837347075343132, 0.018217051401734352, 0.05410189554095268, 0.0016554020112380385, 0.014822332188487053, -0.0005649917293339968, -0.00602276436984539, -0.02949209325015545, ...
0b2bfc773c1b5374373657f87881c3e640608239
subsection
7
160
Tractable models for deep learning
In the following, we assume that is also the case for non-linear networks.In Section REF of the Supplementary Materialwe train a neural network with USV-layers on a simple real-world dataset (MNIST), showing that these layers can learn to represent complex functions despite their restriction. We further note that such ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.053292229771614075, 0.012766022235155106, -0.004864518530666828, -0.0019515303429216146, 0.0011865609558299184, 0.008683642372488976, 0.01405559666454792, 0.041785258799791336, 0.0006548023084178567, 0.01709258183836937, -0.0166500061750412, 0.009553533047437668, -0.017931949347257614, ...
e2564c323564934532963af8d8ea1e45fb78f000
subsection
8
160
Tractable models for deep learning
Additionally, they investigate the training of larger linear networks on i.i.d. normally distributed inputs where entropies at each hidden layer can be computed analytically for an additive Gaussian noise. The strategy proposed in the present paper allows us to evaluate entropies and mutual informations in non-linear n...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.005835222080349922, -0.0020594901870936155, -0.017513293772935867, 0.01609453372657299, -0.012478984892368317, -0.022639136761426926, 0.0240426417440176, -0.006933616939932108, -0.009778764098882675, 0.0296261478215456, -0.03038892149925232, -0.05946587398648262, -0.022776436060667038, ...
57f409eb5236536d09752579813092e500a4911a
subsection
9
160
Numerical experiments
We present a series of experiments both aiming at further validating the replica estimator and leveraging its power in noteworthy applications. A first application presented in the paragraph 3.1 consists in using the replica formula in settings where it is proven to be rigorously exact as a basis of comparison for othe...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.021638013422489166, -0.008110440336167812, -0.0037042510230094194, -0.010292553342878819, 0.008285924792289734, -0.03094632737338543, 0.05075319856405258, -0.009705061092972755, 0.012085547670722008, 0.004204000812023878, -0.010071289725601673, -0.0008936362573876977, -0.04251305013895035...
c857f2f9b94d00ca1a93ae5b52bb6199c1be2644
subsection
10
160
Numerical experiments
To compute entropies, we consider noisy versions of the latent variables where an additive white Gaussian noise of very small variance (\sigma ^2_{\rm noise}=10^{-5}) is added right before the activation function, 1 = f(W_1+ {\epsilon }_1) and 2 = f(W_2 f(W_1 ) + {\epsilon }_2) with {\epsilon }_{1,2} \sim \mathcal {N}(...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.009829262271523476, -0.01975010149180889, -0.009417166002094746, 0.006314996164292097, -0.007291817106306553, -0.038737062364816666, 0.06135657802224159, 0.0019049920374527574, 0.004361353814601898, 0.020879551768302917, -0.055373549461364746, -0.02715257555246353, -0.014850733801722527, ...
7834a4f8db2537ed7bc422b384583ca2de2e4cb6
subsection
11
160
Numerical experiments
This loss of information for bounded activations was also observed by , where entropies were computed by discretizing the output as a single neuron with bins of equal size. In this setting, as the tanh activation starts to saturate for large inputs, the extreme bins (at -1 and 1) concentrate more and more probability m...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.021183323115110397, -0.012507012113928795, -0.029775694012641907, 0.006696097087115049, -0.0023674804251641035, -0.02762378565967083, 0.022602666169404984, 0.036139849573373795, -0.022953687235713005, 0.012865663506090641, -0.03571251779794693, -0.009103639982640743, -0.03546832874417305,...
9fe04f3fda58ccdced5a9866a556cdfc626f24bf
subsection
12
160
Numerical experiments
The teacher has also to be linear to be learnable: we consider a simple single-layer network with additive white Gaussian noise, = \tilde{W}_{\rm teach} + {\epsilon }, with input {x} \sim \mathcal {N}(0, I_{n}) of size n, teacher matrix \tilde{W}_{\rm teach} i.i.d. normally distributed as \mathcal {N}(0, 1/n) , noise {...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.053216706961393356, -0.0032650157809257507, -0.03231450170278549, -0.021298887208104134, 0.028057776391506195, 0.005378121975809336, -0.014784113503992558, 0.03231450170278549, 0.0037131926510483027, 0.02731017954647541, -0.023251794278621674, 0.004054569639265537, -0.03808167949318886, ...
a4fc27e7445111b58a93031a2e6bf3bbad92c3c8
subsection
13
160
Numerical experiments
We then train a recognition model to solve the binary classification problem of recovering the label y = \mathrm {sign}(Y_1), the sign of the first neuron in , using plain SGD but this time to minimize the cross-entropy loss. Note that the rest of the initial code (Y_2, .. Y_{n_{}}) acts as noise/nuisance with respect ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0019879164174199104, -0.021153567358851433, -0.030127808451652527, -0.019047368317842484, 0.015079166740179062, -0.006661999970674515, 0.015132584609091282, 0.017322726547718048, 0.012972966767847538, 0.015353888273239136, -0.04841206222772598, -0.032356105744838715, -0.020375190302729607...
01bfa333538ff73a55d015df015836af658c04c4
subsection
14
160
Conclusion and perspectives
We have presented a class of deep learning models together with a tractable method to compute entropy and mutual information between layers. This, we believe, offers a promising framework for further investigations, and to this aim we provide Python packages that facilitate both the computation of mutual informations a...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03001476265490055, -0.023392288014292717, -0.014252052642405033, 0.011871624737977982, -0.006832287646830082, -0.02890084497630596, 0.017578549683094025, 0.020660899579524994, 0.006885694805532694, 0.004264934454113245, -0.011215480975806713, 0.006542363669723272, -0.018570395186543465, ...
6e8c61894814ca66b68db5bb9b41b7631c472cd4
subsection
15
160
Background
The replica method , was first developed in the context of disordered physical systems where the strength of interactions J are randomly distributed, J \sim P_J (J). Given the distribution of microstates {x} at a fixed temperature \beta ^{-1}, P ({x} | \beta , J) = \frac{1}{\mathcal {Z} (\beta , J)} \, e^{-\beta \mathc...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03225748613476753, 0.016128743067383766, -0.049042366445064545, 0.010833877138793468, 0.002498658373951912, 0.007171721663326025, 0.037323471158742905, 0.004722654819488525, 0.005992965307086706, 0.016479700803756714, -0.05047671124339104, 0.03051796369254589, -0.03131143003702164, 0.02...
67498b35c7c9c697b6373398a5020afc8fe65a3b
subsection
16
160
Background
Note this quantity is nothing but the entropy of {y} given W, H({y} | W).The distribution P_J (or P_W in the notation above) is usually assumed to be i.i.d. on the elements of the matrix J. However, one can also use the same techniques to approach J belonging to arbitrary orthogonally-invariant ensembles. This approach...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05967048183083534, 0.015077861025929451, -0.004578297957777977, -0.0036340239457786083, 0.009812818840146065, -0.028980625793337822, 0.052711471915245056, 0.0167718306183815, 0.02479911409318447, 0.011804378591477871, -0.02698143571615219, 0.0113160265609622, 0.005230705253779888, 0.017...
24263ba79218a7ae675e9bc86608dc0ba4aff4ea
subsection
17
160
Single-layer
For a single-layer generalized linear model\left\lbrace \begin{aligned}&{x} \sim P_X ({x}), \\ &{y} \sim P_{Y | Z} ({y} | W {x}). \end{aligned} \right.with P_X and P_{Y | Z} separable in the components of {x} \in \mathbb {R}^n and {y} \in \mathbb {R}^m, and W \in \mathbb {R}^{m \times n} Gaussian i.i.d., W_{\mu i} \sim...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.030682794749736786, 0.0033127805218100548, -0.005713926628232002, -0.009612212888896465, 0.008948512375354767, -0.036556925624608994, 0.011755889281630516, -0.002197077265009284, -0.00017486503929831088, 0.038876064121723175, -0.0367705300450325, 0.01608138531446457, 0.006213609129190445,...
87cdbd2f0106f3d2e1edc42278016e7ddf1f2235
subsection
18
160
Single-layer
The entropy is then written as n^{-1} H ({y} | W) = \min _{A, V, \tilde{A}, \tilde{V}} \phi (A, V, \tilde{A}, \tilde{V}), where\phi (A, V, \tilde{A}, \tilde{V}) = -\frac{1}{2} \big ( \tilde{A} V + \alpha A \tilde{V} - F_W(AV) \big ) + I(x; x + \frac{\xi _0}{\sqrt{\tilde{A}}}) + \alpha H(y | \xi _1; \tilde{V}, \tilde{\r...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.004318574909120798, -0.002056282479315996, 0.02011265605688095, 0.01732007972896099, 0.0024435045197606087, -0.042300667613744736, 0.046603985130786896, 0.000589416769798845, -0.017091179266572, 0.054447613656520844, -0.03930971398949623, -0.00441394979134202, 0.010063957422971725, 0.00...
80097aec5e32864d0ece4ec4235c9b3fd213f067
subsection
19
160
Multi-layer
Consider the following multi-layer generalized linear model\left\lbrace \begin{aligned}&t_{0,i} \equiv x_i \sim P_0 (x_i), \\ &t_{1,i} \sim P_1 (t_{1,i} | W_1 {x}), \\ &t_{2,i} \sim P_2 (t_{2,i} | W_2 {t}_{1}), \\ &\vdots \\ &t_{L,i} \equiv y_i \sim P_L (y | W_L {t}_{L - 1}), \end{aligned} \right.where the W_\ell \in ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05243720859289169, 0.035436321049928665, -0.024448314681649208, 0.005917499307543039, 0.01680249348282814, 0.028477251529693604, -0.017153499647974968, 0.0396483913064003, -0.010087600909173489, 0.021686052903532982, -0.0364740788936615, -0.0037885732017457485, -0.02078564651310444, 0.0...
764d3cadcff1463fe78d137ebc989532feea44b8
subsection
20
160
Multi-layer
Our analysis employs the framework introduced in to compute the entropy of {t}_L in the limit n_0 \rightarrow \infty with \tilde{\alpha }_\ell = n_\ell / n_0 finite for \ell = 1, \dots , L\lim _{n_0 \rightarrow \infty } n_0^{-1} H({t}_L | {W}) = \min _{{A}, {V}, {\tilde{A}}, {\tilde{V}}} \phi ({A}, {V}, {\tilde{A}}, {...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.06390626728534698, 0.04074253514409065, -0.00974311213940382, 0.011970981024205685, -0.020935865119099617, -0.0030938386917114258, 0.020523861050605774, 0.017410948872566223, 0.01315358281135559, 0.04254314303398132, -0.021241052076220512, 0.019654076546430588, 0.013550326228141785, 0.0...
ffb074c4c99ba015e3cfdd34d30ee5952b6eaa64
subsection
21
160
A simple heuristic derivation of the multi-layer formula
Formula (REF ) can be derived using a simple argument. Consider the case L = 2, where the model reads\left\lbrace \begin{aligned}&{t}_0 \sim P_0 ({t}_0), \\ &{t}_1 \sim P_1 ({t}_1 | W_1 {t}_0), \\ &{t}_2 \sim P_2 ({t}_2 | W_2 {t}_1), \end{aligned} \right.with {t}_\ell \in \mathbb {R}^{n_\ell } and W \in \mathbb {R}^{n_...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0429157093167305, 0.023991286754608154, -0.002022166270762682, -0.024342304095625877, -0.0017684416379779577, -0.017978202551603317, 0.03406396508216858, 0.015948405489325523, 0.05408722534775734, 0.022053059190511703, -0.03873402252793312, -0.012461123056709766, -0.005753635428845882, ...
73aea1c1437ee84d7fe871b9d710d7de57b02ba3
subsection
22
160
A simple heuristic derivation of the multi-layer formula
\end{aligned}Moreover, H(t_1 + \tilde{\xi }_1 / \sqrt{\tilde{A}}_2) can be obtained from the replica free energy of another problem: that of estimating {t}_0 given the knowledge of (noisy) {t}_1, which can again be written using (REF )\lim _{n_0 \rightarrow \infty } n_0^{-1} \, H(t_1 + \frac{\tilde{\xi }_1}{\sqrt{\tild...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0428294911980629, 0.02297162637114525, -0.012577156536281109, 0.0041058920323848724, 0.006567137781530619, 0.014859055168926716, -0.0007598380325362086, 0.011691871099174023, 0.04661484807729721, 0.004193657543510199, -0.027214890345931053, -0.029077041894197464, 0.009005487896502018, 0...
b692d080b33f283bd4b33613da0da63b64c8d6e8
subsection
23
160
Formulation in terms of tractable integrals
While expression (REF ) is more easily written in terms of conditional entropies and mutual informations, evaluating it requires us to explicitely state it in terms of integrals, which we do below. We first consider the Gaussian i.i.d. In this case, the multi-layer formula was derived with the cavity and replica method...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.04204684868454933, 0.03420501947402954, 0.02489856816828251, 0.0001393345301039517, -0.005480888765305281, 0.0005587683990597725, 0.03624938800930977, 0.02904832921922207, -0.003016968024894595, 0.030909620225429535, -0.031886033713817596, 0.024990107864141464, -0.036371439695358276, 0....
9cb4ffcca445586a6a6807e9e7c45202722217a2
subsection
24
160
Formulation in terms of tractable integrals
\end{aligned}where&Z_0 (A, B) \!=\! {\textstyle \int } dx \, P_0 (x) e^{-\frac{1}{2} A x^2 + B x}, \\ &Z_\ell (A, B, V, \omega ) \!=\! {\textstyle \int } dt dz \, P_\ell (t | z) \mathcal {N} (z; \omega , V) e^{-\frac{1}{2} A t^2 + Bt}, \\ &Z_L (y, V, \omega ) \!=\! {\textstyle \int } dz \, P_L(y | z) \mathcal {N} (z; \...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.025691518560051918, 0.038018565624952316, -0.009001185186207294, 0.005099400412291288, 0.0232505202293396, 0.0041725835762917995, 0.046165402978658676, 0.02863597497344017, 0.001217639190144837, 0.03356374055147171, -0.034235015511512756, 0.04341927915811539, 0.01582072675228119, -0.000...
2fc745149be8f45953d4163f0bc323ca814e1f35
subsection
25
160
Recovering the formulation in terms of conditional entropies
One can rewrite the formulas above in a simpler way. By manipulating the measures (REF ) one obtainsK_0 (A, \rho ) = -I(x; b) + \frac{1}{2} A \rho ,for x \sim P_0(x) and b \sim \mathcal {N} (b; Ax, A). Introducing a standard normal variable \xi _0 and using the invariance of mutual informations, this can be written asK...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03965349867939949, 0.03124353662133217, -0.028862494975328445, -0.000927232438698411, -0.00047410960542038083, 0.00005914445864618756, 0.03534930571913719, 0.00018482643645256758, 0.007066808640956879, 0.03244931995868683, -0.03406720608472824, 0.02434462122619152, 0.003922612871974707, ...
160769964679f2b37ae02c4de3e7a105e63b7db6
subsection
26
160
Recovering the formulation in terms of conditional entropies
Introducing standard normal \xi _\ellK_\ell (A, V, \rho ) = -H(t_\ell | \xi _\ell ; A, V) + \frac{1}{2} A \rho + \frac{1}{2} \log (2 \pi e A^{-1})\,,whereP(t_\ell | \xi _\ell ; A, V, \rho ) = \int \mathcal {D}\tilde{\xi } \mathcal {D}\tilde{z} \, P_\ell (t_\ell + \sqrt{1 / A} \tilde{\xi } | \sqrt{\rho - V} \xi _\ell + ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05858412757515907, 0.0047485181130468845, -0.032678958028554916, -0.0034669903106987476, -0.02198430337011814, 0.013799307867884636, 0.002395236399024725, 0.019100865349173546, 0.016186915338039398, -0.002210254082456231, -0.03136691823601723, 0.03359433636069298, -0.009603830054402351, ...
c00f4229f795864537552f4749019bae15e96329
subsection
27
160
Solving saddle-point equations
In order to deal with the extremization problem in\lim _{n_0 \rightarrow \infty } n_0^{-1} H ({t}_L | {W}) = \min _{{A}, {V}, \tilde{{A}}, \tilde{{V}}} \phi ({A}, {V}, \tilde{{A}}, \tilde{{V}}),one needs to solve the saddle-point equations \nabla _{\lbrace {A}, {V}, \tilde{{A}}, \tilde{{V}}\rbrace } \phi = 0. In what f...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03933761641383171, -0.0043831379152834415, 0.028366418555378914, 0.0038452595472335815, -0.034668371081352234, 0.036224786192178726, 0.029892316088080406, 0.01915000192821026, 0.011116157285869122, 0.009285081177949905, -0.053009647876024246, -0.022217055782675743, 0.006164622493088245, ...
58f255ff22fcbe8bd68316aedfd550aec0d58f62
subsection
28
160
Method 1: fixed-point iteration
We first introduce the following function, which is related to the derivatives of F_{W_\ell }\psi _\ell (\theta , \gamma ) = 1 - \gamma \big [ \mathcal {S}_\ell \big ( -\gamma ^{-1} (1 - \theta ) (1 - \alpha _\ell \theta ) \big ) \big ]^{-1},where \mathcal {S}_\ell (z) = \mathbb {E}_{\lambda _\ell } \frac{1}{\lambda _\...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.038539353758096695, 0.06401865929365158, -0.019010307267308235, -0.001384578994475305, 0.012312454171478748, -0.015836836770176888, 0.060265421867370605, 0.021832864731550217, 0.007071651052683592, 0.02639472857117653, 0.0003213520103599876, 0.03710519149899483, -0.0068504237569868565, ...
265d46a87f1b866fc5db22a0f94dc9854ec31dc6
subsection
29
160
Method 1: fixed-point iteration
After these quantities are computed for all layers, we compute all the V_\ell ; for 2 \le \ell \le LV_\ell ^{(t + 1)} = \mathbb {E}_{b, t, z, w | \tilde{\rho }_{\ell - 1}, \tilde{A}_{\ell }, \tilde{V}_{\ell - 1}} \partial _b^2 \log Z_\ell (\tilde{A}_{\ell }^{(t)}, b, \tilde{V}_{\ell - 1}^{(t)}, w),and for the 1st layer...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.015248758718371391, 0.023449067026376724, -0.02521880716085434, -0.007670148275792599, -0.005171915050595999, -0.012281390838325024, 0.027873417362570763, 0.014783438295125961, 0.013967221602797508, 0.0338386669754982, 0.006750951055437326, 0.01276196725666523, -0.009786971844732761, 0....
d59bcfc85ce72119bb2750a5e9c6c82534f0d6ab
subsection
30
160
Method 2: ML-VAMP state evolution
While the fixed-point iteration above works well in most cases, it is not provably convergent. In particular, it relies on a solution for \theta = \psi _\ell (\theta , A_\ell ^{(t)} V_\ell ^{(t)}) being found, which might not happen throughout the iteration.An alternative is to employ the state evolution (SE) of the ML...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.012508774176239967, 0.025048058480024338, -0.024224309250712395, 0.025566713884472847, -0.012783356942236423, 0.0023282337933778763, 0.01615462452173233, 0.0573878176510334, 0.004404766485095024, 0.018519088625907898, 0.0018858503317460418, -0.0016436834121122956, 0.010746867395937443, ...
5530dee7d062887f7fd4ac88354622015e6b3650
subsection
31
160
Method 2: ML-VAMP state evolution
Let us first look at the single-layer case; the ML-VAMP SE equations read&A^+_x = \frac{1}{V^+_x (A^-_x)} - A^-_x, &&\qquad A^+_z = \frac{1}{V^+_z (A^+_x, 1 / A_z^-)} - A^-_z, \\ &A^-_x = \frac{1}{V^-_x (A^+_x, 1 / A_z^-)} - A^+_x, &&\qquad A^-_z = \frac{1}{V^-_z (A^+_z)} - A^+_z,where&V^+_x (A) = \mathbb {E}_{x, z} \p...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.02731863595545292, 0.050241872668266296, -0.012041566893458366, -0.003775389865040779, 0.019809827208518982, 0.024556249380111694, 0.027852799743413925, 0.03861237317323685, -0.019626686349511147, 0.011141120456159115, -0.002848234958946705, 0.013010691851377487, -0.02228224091231823, 0...
be074e7e47cc147e1fc4691d8e7190240ba5ad51
subsection
32
160
Mutual information from entropy
While in our computations we focus on the entropy H(\ell ), the mutual information I(\ell ; {\ell - 1}) can be easily obtained from the chain rule relationI(\ell ; {\ell -1}) &= H(\ell ) + {\ell }{\ell -1} \log P_{\ell | {\ell - 1}} ({t}_\ell | {t}_{\ell -1}) \\ &= H(\ell ) + \int dz \, \mathcal {N} (z; 0, \tilde{\rho ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.01097586564719677, 0.029701197519898415, -0.020029621198773384, 0.009183420799672604, -0.029487628489732742, -0.05516154319047928, 0.020426247268915176, 0.04018127918243408, -0.02356874570250511, -0.011525040492415428, -0.028221476823091507, -0.014415835030376911, 0.008435932919383049, ...
b29981b3cf8d7ed25260a9c87d25df4e0c9e6138
subsection
33
160
Equivalence in linear case
In the linear case, = W_L W_{L - 1} \cdots W_1 + \mathcal {N} (0, \Delta ), our formula reduces to , ,\lim _{n_ \rightarrow \infty } \, n^{-1} I(; ) = \min _{A, V} \, \left\lbrace -\frac{1}{2} A V - \frac{1}{2} G(-V / \Delta ) + I (x; x + \sqrt{1 / A} \xi ) \right\rbrace ,whereG(x) = _{\Lambda } \left\lbrace -\mathbb ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.048722416162490845, 0.0533292330801487, 0.01229501236230135, 0.0033140091691166162, -0.010739067569375038, -0.001544503727927804, 0.007855993695557117, 0.025306979194283485, 0.003523756517097354, 0.03725114092230797, -0.02951718121767044, -0.00656700087711215, 0.0076691280119121075, 0.0...
e5c2645510904dc5a0f49696f3d69b4f7d577b31
subsection
34
160
Proof of the replica formula by the adaptive interpolation method
In this section we prove Theorem (that we re-write below more explicitely) using the adaptive interpolation method of , in a multi-layer setting, as first developed in .
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ 0.019468121230602264, 0.027829037979245186, 0.009421288967132568, 0.02341972291469574, -0.004470344632863998, -0.03173486888408661, -0.0030247296672314405, 0.024731837213039398, 0.0049471305683255196, 0.0038181014824658632, -0.005984616465866566, -0.003751351498067379, -0.013967919163405895,...
0bba10ba354c1f3bc9f40896386dbd47dee28f9b
subsection
35
160
Two-layer generalized linear estimation: Problem statement
One gives here a generic description of the observation model, that is a two-layer generalized linear model (GLM). Let n_0, n_1, n_2 \in ^* and define the triplet \mathbf {n} = (n_0,n_1,n_2). Let P_0 be a probability distribution over and let (X^0_i)_{i=1}^{n_0} P_0 be the components of a signal vector ^0. One fixes tw...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05554515868425369, -0.023103123530745506, -0.03108392469584942, -0.038912128657102585, 0.015076542273163795, -0.00723689328879118, 0.003416256047785282, 0.0025216585490852594, 0.009705143049359322, 0.046114686876535416, -0.05878020450472832, 0.011887273751199245, 0.007122445851564407, 0...
a19b77653aa0f99f66e43d481e196fb2e60136e1
subsection
36
160
Two-layer generalized linear estimation: Problem statement
The estimation problem is to recover ^0 from the knowledge of =(Y_\mu )_{\mu =1}^{n_2}, \varphi _1, \varphi _2, {1}, {2} and \Delta , P_0.In the language of statistical mechanics, the random variables , {1}, {2}, ^0, _1, _2, are called quenched variables because once the measurements are acquired they have a “fixed rea...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.021127674728631973, -0.004976826254278421, -0.009457875974476337, -0.023537907749414444, -0.02216499112546444, -0.043323174118995667, 0.012142692692577839, 0.02527693659067154, 0.027549877762794495, 0.017039431259036064, -0.053208183497190475, 0.0055145518854260445, -0.013012207113206387,...
77bc32012de8b140da8d5512931aede1dbdeb527
subsection
37
160
Two-layer generalized linear estimation: Problem statement
The partition function is defined as(,{1}, {2}) \\ \!\int \!\! dP_0()dP_{A_1}(_1)dP_{A_2}(_2) \prod _{\mu =1}^{n_2} \frac{1}{\sqrt{2\pi \Delta }}e^{-\frac{1}{2 \Delta }\Big ( Y_\mu - \varphi _2\Big (\frac{1}{\sqrt{n_1}} \Big [{2} \varphi _1\Big (\frac{{1} }{\sqrt{n_0}},_1\Big )\Big ]_{\mu },_{2,\mu }\Big )\Big )^2}.One...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03164321184158325, -0.032345037907361984, -0.02944619208574295, 0.003131518606096506, -0.009726397693157196, -0.018491597846150398, 0.022778840735554695, -0.010946964845061302, 0.00009112538100453094, 0.03164321184158325, -0.03231452405452728, 0.005942638032138348, 0.010077310726046562, ...
0e1a8eafc0c20590af0949f3786f70f6d2cf7b14
subsection
38
160
Important scalar inference channels
The thermodynamic limit of the free entropy will be expressed in terms of the free entropy of simple scalar inference channels. This “decoupling property” results from the mean-field approach in statistical physics, used through in the replica method to perform a formal calculation of the free entropy of the model , . ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.04193710535764694, -0.006031892728060484, -0.023242436349391937, 0.012628440745174885, 0.015398303978145123, -0.010453755035996437, 0.023746047168970108, 0.003746564732864499, 0.01912197843194008, 0.02322717383503914, -0.01962558925151825, 0.011735674925148487, -0.002449383493512869, -0...
296061db4d2fd12b23a5e99780060f662afc850d
subsection
39
160
Important scalar inference channels
Suppose that V,U (0,1), where V is known. Consider the problem of recovering the unknown U from the observation Y_0^{\prime } = \sqrt{r}\varphi _1(\sqrt{q}\, V + \sqrt{\rho - q} \,U,_1) + Z^{\prime } where r \ge 0, \rho >0, q \in [0, \rho ], Z^{\prime } \sim (0,1) and _1 \sim P_{A_1}. Equivalently, Y_0^{\prime } \sim P...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.027410367503762245, 0.008684019558131695, -0.023945916444063187, -0.003071456914767623, 0.013957004994153976, 0.016971230506896973, 0.03461398184299469, 0.034797124564647675, 0.05021164193749428, 0.044534217566251755, -0.026021534577012062, -0.009470007382333279, 0.00428096204996109, 0....
e497a27129dd2b0081be099adacdac03719640ee
subsection
40
160
Important scalar inference channels
Hence\Psi _{P_{\rm out,1}^{(r)}}(q;\rho ) = -\frac{1 + \ln (2\pi ) + r [\varphi _1^2(T,_1)]}{2} + \Psi _{\varphi _1}(q, r;\rho )where\Psi _{\varphi _1}(q, r;\rho ) \ln \int \! {\cal D}u dP_{A_1}() dh e^{\sqrt{r}h Y_0^{\prime } -\frac{r h^2}{2}} \delta \big (h -\varphi _1(\sqrt{q}\, V + \sqrt{\rho - q}\, u, )\big ) \; .
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.017733518034219742, 0.02678341045975685, -0.0035425066016614437, 0.006547055672854185, -0.0057496577501297, -0.0023807475809007883, 0.004322735592722893, 0.015352769754827023, 0.002815691754221916, 0.02005321905016899, -0.019671689718961716, 0.02051105536520481, -0.01974799670279026, 0....
aed990e014587cc74d043981c68414bfee583f61
subsection
41
160
Replica-symmetric formula and mutual information
Our goal is to prove Theorem  that gives a single-letter replica-symmetric formula for the asymptotic free entropy of model (REF ), (REF ). The result holds under the following hypotheses:The prior distribution P_0 has a bounded support. \varphi _1, \varphi _2 are bounded \mathcal {C}^2 functions with bounded first a...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05768365040421486, 0.02781939134001732, 0.017900243401527405, -0.021242232993245125, 0.013436628505587578, -0.036929745227098465, 0.034884873777627945, 0.020784426480531693, 0.04062271863222122, 0.001382004120387137, -0.023088721558451653, 0.018617475405335426, -0.017396656796336174, -0...
4e5158d889bb0a3394fe2190bfa57de0570ee289
subsection
42
160
Interpolating estimation problem
The proof of Theorem REF follows the same steps than the proof of the replica formula for a one-layer GLM in . Let t\in [0,1] be an interpolation parameter. We introduce an interpolating estimation problem that interpolates between the original problem (REF ) at t=0 and two analytically tractable problems at t=1.Prior ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03305619955062866, 0.017779534682631493, 0.0036684595979750156, -0.023243116214871407, 0.017046986147761345, 0.01652809977531433, -0.011140823364257812, 0.030034439638257027, 0.01755061373114586, 0.02927136979997158, -0.029881825670599937, 0.015398754738271236, -0.030324406921863556, 0....
b911d18b6054c3b8a6d0a22c9312cf420d6aa0a9
subsection
43
160
Interpolating estimation problem
Let q_{\epsilon }: [0,1] \rightarrow [0,\rho _1(n_0)] and r_{\epsilon }: [0,1] \rightarrow [0,] be two continuous “interpolation functions”. Their dependence on \epsilon will also be specified later. It is useful to defineR_1(t,\epsilon ) \epsilon _1 + \int _0^t r_{\epsilon }(v)dv\,,\qquad R_2(t,\epsilon ) \epsilon _2 ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.01494151633232832, 0.04215369373559952, -0.024922875687479973, 0.015529103577136993, 0.012110411189496517, -0.011095487512648106, 0.019855886697769165, 0.0040864055044949055, 0.006848830729722977, 0.000052284336561569944, 0.008500943891704082, -0.02104632370173931, 0.022053616121411324, ...
39cc1dbeed14f7754017ff1884a93b9af7da524c
subsection
44
160
Interpolating estimation problem
The inference problem is to estimate both unknowns ^0 and from the knowledge of , {1}, {2} and the two kinds of observations{\left\lbrace \begin{array}{ll} Y_{t,\epsilon ,\mu } &\sim P_{\rm out, 2}(\ \cdot \ | \, S_{t,\epsilon ,\mu })\,, \qquad \qquad \qquad \qquad \;\;\, 1 \le \mu \le n_2, \\ Y^{\prime }_{t,\epsilon ,...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ 0.002700991230085492, 0.012856412678956985, 0.0034525380469858646, 0.0318014994263649, 0.005249383859336376, -0.03277812898159027, -0.008728627115488052, 0.007629918400198221, 0.0387294664978981, 0.025468667969107628, -0.043551575392484665, -0.008736256510019302, -0.017838750034570694, 0.0...
dc88af3d5cd6139bbbcc33970b98f02d76b777d1
subsection
45
160
Interpolating estimation problem
When the (t,\epsilon )-dependent observations (REF ) are considered, it reads_{t,\epsilon }(,_1, ;_{t,\epsilon },_{t,\epsilon }^{^{\prime }},{1},{2},) - \sum _{\mu =1}^{n_2} \ln P_{\rm out,2} ( Y_{t,\epsilon ,\mu } |s_{t, \epsilon ,\mu })\\ + \frac{1}{2} \sum _{i=1}^{n_1}\Bigg [ \sqrt{R_1(t,\epsilon )} \bigg ( \varphi ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0189930722117424, -0.010503397323191166, -0.017055625095963478, -0.03371460735797882, 0.01305869035422802, -0.023920591920614243, -0.006147958338260651, -0.019603291526436806, 0.03987782448530197, 0.01501901913434267, -0.0401829332113266, -0.013546865433454514, 0.014538471587002277, 0.0...
0b309009800b4a1a44225301fa8fdf6ce8c1e144
subsection
46
160
Interpolating estimation problem
\mathbf {n} and the choice of (q_{\epsilon })_{\epsilon \in {\cal B}_{n_0}}, (r_{\epsilon })_{\epsilon \in {\cal B}_{n_0}}. A simple computation shows that \frac{\partial f_{\mathbf {n},\epsilon }(0)}{\partial \epsilon _1} = - \frac{n_1}{n_0} \langle \mathcal {L} \rangle _{\mathbf {n},0,\epsilon }, where\mathcal {L} ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.016635386273264885, 0.05741497874259949, -0.027944397181272507, -0.03623156622052193, 0.005143234506249428, -0.0025811558589339256, 0.015658630058169365, 0.006928867660462856, 0.010553550906479359, 0.037727225571870804, -0.04810526594519615, 0.006253532133996487, -0.023182708770036697, ...
31cd3409f7fac5731ab47601a2007d2c09c00236
subsection
47
160
Interpolating estimation problem
In a similar fashion to what is done in Appendix REF , we compute\Big \vert \frac{\partial f_{\mathbf {n},\epsilon }(0)}{\partial {\epsilon _2}}\Big \vert = \frac{1}{2 n_0}\sum _{\mu =1}^{n_2} \big \vert \big [u^{\prime }_{Y_{0,\epsilon ,\mu }}(S_{0,\epsilon ,\mu })\langle u^{\prime }_{Y_{0,\epsilon ,\mu }}(s_{0,\epsil...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05655229091644287, 0.05993993580341339, -0.021546635776758194, -0.01864730194211006, -0.0013991198502480984, 0.016739845275878906, 0.009445727802813053, 0.01470268052071333, 0.017243413254618645, 0.0191966500133276, -0.03497513383626938, 0.015122320502996445, -0.023972922936081886, 0.00...
a7e04dccb64c38f12305cbd8bb47ecf234c3d711
subsection
48
160
Interpolating free entropy at t=0 and t=1
We will denote by _{n_0}(1) any quantity that vanishes uniformly in t \in [0,1] and \epsilon when n_0 \rightarrow +\infty . It is easily shown (the first equality uses Lemma REF ) that\left\lbrace \begin{array}{lll} f_{\mathbf {n},\epsilon }(0) &=& f_{\mathbf {n},\epsilon =(0,0)}(0) + _{n_0}(1) = f_{\mathbf {n}} - \fr...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.05610769987106323, 0.05073503404855728, -0.0109819071367383, -0.011455067433416843, 0.01607220061123371, -0.024192247539758682, -0.009867689572274685, 0.018086949363350868, 0.0480792298913002, 0.04380552098155022, -0.008921368047595024, -0.006521222647279501, -0.0244669858366251, 0.0103...
20b434027ba632792265af0fed446b66e799d8f2
subsection
49
160
Interpolating free entropy at t=0 and t=1
Applying Theorem 1 of , then (REF ), the free entropy \tilde{f}_{(n_0,n_1),\epsilon } in the thermodynamic limit n_0, n_1 \rightarrow +\infty such that {n_1}{n_0} \rightarrow \alpha _1 satisfies\tilde{f}_{n_0,n_1} &= _{n_0}(1) + {\sup }_{q_0 \in [0,\rho _0]} {\inf }_{r_0 \ge 0} \Big \lbrace \psi _{P_0}(r_0) + \alpha _1...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.015950286760926247, 0.04807981103658676, -0.008929107338190079, 0.0026653767563402653, 0.008974897675216198, -0.02523045241832733, 0.02666521444916725, -0.029321052134037018, 0.032694268971681595, 0.033732183277606964, -0.00022740128042642027, 0.0066701197065413, -0.007536319550126791, ...
5c46385bfc1655257463024fd3191df77c9f068f
subsection
50
160
Interpolating free entropy at t=0 and t=1
The second line follows from the Lipschitzianity in both its arguments of the continuous mapping (q,\rho ) \mapsto \Psi _{P_{\rm out},2}\big (q;\rho \big ) on the compact \big \lbrace (q,\rho ): 0 \le \rho \le 1 + \rho _{\mathrm {max}}, 0 \le q \le \rho \big \rbrace , with \rho _{\mathrm {max}} a upper bound on the seq...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.027887441217899323, 0.01910015195608139, 0.011136669665575027, 0.007833807729184628, 0.006197632756084204, -0.039939455687999725, 0.025034623220562935, -0.002357007469981909, -0.001417827676050365, -0.0011870850576087832, -0.0018068483332172036, 0.004115228075534105, -0.010732393711805344...
52782e4ad9b7f023f80c6de01fcdba5ab4146d20
subsection
51
160
Interpolating free entropy at t=0 and t=1
Reporting (REF ) and (REF ) in (REF ), we obtain:f_{\mathbf {n},\epsilon }(1) = -\frac{\alpha _1}{2}\bigg (1 + \rho _1 \int _0^1 \!\! r_{\epsilon }(t) dt\bigg ) + \alpha \Psi _{P_{\rm out, 2}}\bigg (\int _0^1 q_{\epsilon }(t) dt;\rho _1(n_0)\bigg )\\ + {\sup }_{q_0 \in [0,\rho _0]} {\inf }_{r_0 \ge 0} \; \bigg \lbrace ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0423540398478508, 0.005027253646403551, -0.015493706800043583, -0.009093303233385086, 0.039668772369623184, -0.008345698937773705, -0.007964268326759338, 0.010626653209328651, 0.01717962883412838, 0.03557983413338661, -0.02070404589176178, 0.03795996308326721, 0.0017479051603004336, 0.0...
7d3086f3237c55ba5c9e9216616d142061493ec9
subsection
52
160
Free entropy variation along the interpolation path
From the Fundamental Theorem of Calculus and (\ref {eq:f0_f1}), (\ref {eq:f1_thermoLimit})f_{\mathbf {n}} &= f_{\mathbf {n},\epsilon }(0) + \frac{1}{2}\frac{n_1}{n_0} + _{n_0}(1)\\ &= f_{\mathbf {n},\epsilon }(1) - \int _0^1\frac{df_{\mathbf {n},\epsilon }(t)}{dt} dt + \frac{1}{2}\alpha _1 + _{n_0}(1)\\ &= -\frac{\alph...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ 0.013868162408471107, 0.04674592986702919, -0.018078958615660667, 0.04561695083975792, 0.04637977480888367, -0.012327255681157112, 0.02877376787364483, 0.002168330829590559, 0.02004704810678959, 0.03313712775707245, 0.0047142598778009415, -0.0064687579870224, -0.035212013870477676, -0.0115...
d8602c71f33fdaaf3df9ef74bf1d1bc6ffb47e68
subsection
53
160
Free entropy variation along the interpolation path
The overlap \widehat{Q}: \widehat{Q} \frac{1}{n_1}\sum _{i=1}^{n_1} \varphi _1\bigg (\bigg [\frac{{1} }{\sqrt{n_0}}\bigg ]_i,_{1,i}\bigg ) \varphi _1\bigg (\bigg [\frac{{1} ^0}{\sqrt{n_0}}\bigg ]_i, _{1,i}\bigg )\,.In Appendix we show [Free entropy variation] The derivative of the free entropy (REF ) verifies, for al...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.014239930547773838, 0.03391331806778908, -0.027106232941150665, -0.03812577202916145, 0.010584556497633457, -0.013125766068696976, -0.009691699407994747, 0.0004810078244190663, 0.023946892470121384, 0.0007526329718530178, -0.014499393291771412, 0.007139043416827917, -0.032356541603803635,...
b0e6388f6a017257252c67e933262c10ef3f904a
subsection
54
160
Overlap concentration
An important quantity appearing naturally in the t-derivative of the average free entropy is \widehat{Q}, the overlap between the hidden output ^1 and the sample ^1 \varphi _{1}\big ({{1} }{\sqrt{n_0}}, _{1}\big ) where the triplet (,,_1) is sampled from the posterior distribution associated to the Gibbs bracket \langl...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.015034439973533154, 0.0046209916472435, -0.03029783070087433, -0.012928091920912266, 0.01863659918308258, -0.020040832459926605, 0.022681398317217827, 0.0026615536771714687, 0.02545933611690998, -0.0010484041413292289, -0.026527773588895798, 0.0009630245622247458, -0.03269418329000473, ...
f3f8c24a36a5605d6363e2e4e86c465ceab632b9
subsection
55
160
Overlap concentration
For now we assume that we can take q_\epsilon (t) = \big [\langle \widehat{Q} \rangle _{\mathbf {n},t,\epsilon }\big ] and prove Assume that  REF , REF , REF hold, that the interpolation functions (q_{\epsilon })_{\epsilon \in {\cal B}_{n_0}}, (r_{\epsilon })_{\epsilon \in {\cal B}_{n_0}} are regular, and that \forall...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.005862188059836626, 0.048148512840270996, -0.007746326737105846, -0.0141729936003685, 0.012014243751764297, -0.008711280301213264, 0.014775613322854042, -0.023036835715174675, 0.02665255218744278, 0.03478409722447395, 0.00019463551871012896, -0.007273385301232338, 0.02198415994644165, 0...
eecbbcbc4358c971caf9549127ab124ebb12582d
subsection
56
160
Overlap concentration
2}\\ \le \frac{1}{s_{n_0}^2} \int _{{\cal B}_{n_0}} \!\!\!\! d\epsilon \int _0^1 \!\!dt\, \Bigg [\Bigg \langle \Bigg (\frac{1}{n_1}\sum _{\mu =1}^{n_2} u_{Y_{t,\epsilon ,\mu }}^{\prime }( S_{t,\epsilon ,\mu } )u_{Y_{t,\epsilon ,\mu }}^{\prime }(s_{t,\epsilon ,\mu }) - r_{\epsilon }(t)\Bigg )^{\! 2} \Bigg \rangle _{\! \...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.02345351316034794, 0.06161696836352348, -0.047212209552526474, -0.020615287125110626, -0.006683564279228449, 0.005829044617712498, -0.01049838401377201, 0.01818906143307686, 0.004890598822385073, 0.005703155416995287, -0.028946852311491966, 0.010818828828632832, 0.015625502914190292, 0....
1e5b2a51dfb7fde9076beb7ab0c72f06dee822c7
subsection
57
160
Overlap concentration
2} \Bigg \rangle _{\! \mathbf {n},t,\epsilon }\Bigg ]\\ &\qquad \qquad \qquad \qquad \qquad \qquad \le \frac{1}{n_1^2}\Big [ \Big \Vert \big \lbrace u_{Y_{t,\epsilon ,\mu }}^{\prime }( S_{t,\epsilon ,\mu })\big \rbrace _{\mu =1}^{n_2} \Big \Vert \cdot \Big \langle \Big \Vert \big \lbrace u_{Y_{t,\epsilon ,\mu }}^{\prim...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.027779048308730125, 0.0509486049413681, -0.04267594590783119, -0.024085350334644318, -0.011355830356478691, 0.012241097167134285, -0.020406916737556458, 0.010035562328994274, 0.031533800065517426, -0.012111359275877476, 0.0008218286093324423, 0.003418959677219391, -0.005075018387287855, ...
9ace7ff9c3bed9142fe27fda8b9e37d3ba85d116
subsection
58
160
Overlap concentration
\mathbf {n},t,\epsilon }\Big ]\\ &\qquad \qquad \qquad \qquad \qquad \qquad \le \frac{1}{n_1^2}\Big [ \Big \Vert \big \lbrace u_{Y_{t,\epsilon ,\mu }}^{\prime }(S_{t,\epsilon ,\mu })\big \rbrace _{\mu =1}^{n_2}\Big \Vert ^2 \Big ] = \frac{n_2}{n_1^2}\Big [ u_{Y_{t,\epsilon ,1}}^{\prime }(S_{t,\epsilon ,1})^2\Big ]\,.Th...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.02110101655125618, 0.03570237010717392, -0.024564450606703758, -0.012183052487671375, -0.0243508480489254, 0.038296133279800415, 0.027112441137433052, 0.00840684026479721, 0.04131710156798363, -0.003438641782850027, -0.013907141052186489, 0.0008901752880774438, -0.006865840405225754, 0....
683a193df55fca2356b1f0755ad2430c5aac0862
subsection
59
160
Overlap concentration
Putting everything together, we obtain the following bound uniform w.r.t. the choice of the interpolating functions:\frac{1}{s_{n_0}^2} \int _{{\cal B}_{n_0}} \!\!\!\! d\epsilon \int _0^1 \!\!dt\, \Bigg [\Bigg \langle \Bigg (\frac{1}{n_1}\sum _{\mu =1}^{n_2} u_{Y_{t,\epsilon ,\mu }}^{\prime }( S_{t,\epsilon ,\mu } )u_{...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.040989529341459274, 0.052221208810806274, -0.03073451668024063, -0.008675557561218739, -0.022417578846216202, -0.01817517913877964, -0.016878042370080948, 0.00037626511766575277, 0.02426409162580967, 0.020265858620405197, -0.020174294710159302, 0.006378862075507641, 0.03506848216056824, ...
02738cdd6bb4fb8a33fbd14021dfab1cbd1df1d2
subsection
60
160
Overlap concentration
\mathbf {n},t,\epsilon }\,\Bigg ]\Bigg \vert \\ \le \frac{C(\varphi _1,\varphi _2,\alpha _1,\alpha _2, S)}{n^{{1}{16}}}\,.Therefore the integral of (REF ) over (t,\epsilon ) reads\frac{1}{s_{n_0}^2}\int _{{\cal B}_{n_0}} \!\!\!\! d\epsilon \int _0^1 \!\! dt\, \frac{df_{\mathbf {n},\epsilon }(t)}{dt} &= \frac{1}{s_{n_0}...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ 0.009252323769032955, 0.061539772897958755, -0.041250258684158325, -0.014309446327388287, -0.015697676688432693, 0.0035754546988755465, 0.019358942285180092, -0.009900672361254692, 0.03459896147251129, -0.012783919461071491, 0.019847111776471138, 0.007757306564599276, -0.01125076413154602, ...
5080d33024c51340d2fc5895ce6d084eca3b02ed
subsection
61
160
Lower and upper matching bounds
To end the proof of Theorem REF one has to go through the following two steps:Prove that under the assumptions REF , REF and REF \lim _{\mathbf {n} \rightarrow \infty }f_{\mathbf {n}} = {\sup }_{r_1 \ge 0} {\inf }_{q_1 \in [0,\rho _1]} {\sup }_{q_0 \in [0,\rho _0]} {\inf }_{r_0 \ge 0} f_{\rm RS}(q_0,r_0,q_1,r_1;\rho _...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.02579815872013569, 0.02961219847202301, 0.007311512716114521, -0.02569136582314968, -0.008490050211548805, 0.025782903656363487, 0.004298422019928694, 0.005129882134497166, 0.0033163067419081926, 0.02498958259820938, -0.025630341842770576, -0.006460981909185648, -0.016232550144195557, 0...
945e7b482c0484529ec92644fc22c92f3124ee26
subsection
62
160
Lower and upper matching bounds
Fix (t,r_1,r_2) \in D_{\mathbf {n}}^{\circ }. Consider the two sets of observations{\left\lbrace \begin{array}{ll} Y_{t,r_2,\mu } &\sim P_{\rm out, 2}(\ \cdot \ | \, S_{t,r_2,\mu })\,, \qquad \qquad \;\;\; 1 \le \mu \le n_2, \\ Y^{\prime }_{t,r_1,i} &= \sqrt{r_1}\, \varphi _1\Big (\Big [\frac{{1} ^0}{\sqrt{n_0}}\Big ]_...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0012837464455515146, 0.011673890054225922, -0.0022851829417049885, -0.00470389099791646, -0.004272796213626862, 0.0003347659658174962, -0.027330636978149414, 0.016557086259126663, 0.01947174407541752, 0.024171819910407066, -0.026613418012857437, -0.01698436588048935, -0.001260856399312615...
03f519020095b426dc144e4ebaad172666c26a10
subsection
63
160
Lower and upper matching bounds
The Gibbs bracket corresponding to this Hamiltonian is denoted \langle - \rangle _{\mathbf {n},t,r_1,r_2} and is used to define F_{\mathbf {n}}:F_{\mathbf {n}}(t,r_1,r_2) \big [\langle \widehat{Q} \rangle _{\mathbf {n},t,r_1,r_2}\big ] = \frac{1}{n_1} \sum _{i=1}^{n_1} \big [X_i^1 \langle x_i^1 \rangle _{\mathbf {n},t,...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.0363384373486042, -0.007887057960033417, 0.0023207326885312796, -0.008192166686058044, 0.010373693890869617, -0.016002947464585304, 0.006201332900673151, 0.04787154495716095, 0.0024465899914503098, 0.03279917687177658, -0.050007306039333344, -0.004736811388283968, -0.0118382154032588, 0...
f11dfc2cbc2149d33c90e019bff99322b1392df2
subsection
64
160
Lower and upper matching bounds
Then \mathrm {MMSE}\big ( ^1 \, \big \vert \, _{t,r_2}, \, _{t,r_1}^{\prime },{1},{2},\big ) is clearly a non-increasing function of r_1, so F_{\mathbf {n}}(t,r_1,r_2) is a non-decreasing function of r_1.r_2's only role is in the generation process of _{t,r_2}:_{t,r_2} \sim P_{\rm out}\Big ( \cdot \, \Big \vert \, \sqr...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.024056674912571907, 0.0006464182515628636, -0.007146830670535564, 0.021204045042395592, 0.009572329930961132, 0.022927826270461082, -0.004332337062805891, 0.03313322737812996, 0.0038422795478254557, 0.024514315649867058, -0.025841476395726204, -0.03148571774363518, -0.03298068046569824, ...
b7212eb52a7b8fc5220792949f5b70577a9a3444
subsection
65
160
Lower and upper matching bounds
Now notice that\mathrm {MMSE}\big ( ^1 \, \big \vert \, _{t,r_2}, \, _{t,r_1}^{\prime },{1},{2},\big ) &\ge \mathrm {MMSE}\big ( ^1 \, \big \vert \, \widetilde{}_{t,r_2,r_2^{\prime }}, _{t,r_1}^{\prime },{1},{2},\big )\,;\\ \mathrm {MMSE}\big ( ^1 \, \big \vert \, _{t,r_2^{\prime }}, \, _{t,r_1}^{\prime },{1},{2},\big ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.04953405261039734, 0.017198052257299423, 0.005497425328940153, -0.006817417684942484, 0.022478021681308746, 0.002872700337320566, 0.040408555418252945, 0.02504170686006546, 0.0016948169795796275, 0.04141571745276451, -0.017671113833785057, -0.031145717948675156, -0.009758788160979748, 0...
1b52760847ce5bd923f1a9d0451328aa57a1465e
subsection
66
160
Lower and upper matching bounds
Define for t \in [0,1]r_{\epsilon }(t) R_1^{\prime }(t,\epsilon ) = r \qquad \text{and} \qquad q_{\epsilon }(t) R_2^{\prime }(t,\epsilon ) = F_{\mathbf {n}}(t,R(t,\epsilon )) \in [0,\rho _1(n_0)].Clearly R_1(t,\epsilon ) = \epsilon _1 + \int _0^t r_\epsilon (s) ds and R_2(t,\epsilon ) = \epsilon _2 + \int _0^t q_\epsil...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.03415504842996597, 0.046913594007492065, -0.04798189178109169, -0.0014021425740793347, 0.03598641976714134, -0.006909606046974659, -0.013056141324341297, 0.008569285273551941, 0.02562391757965088, 0.02939348667860031, 0.006585300900042057, -0.011552891694009304, -0.0057268464006483555, ...
5d2b004930dd408027ca6597b998578191ddedf7
subsection
67
160
Lower and upper matching bounds
Moreover, since also the Jacobian determinant is greater than or equal to 1 the functions (q_{\epsilon })_{\epsilon \in {\cal B}_{n_0}} and (r_{\epsilon })_{\epsilon \in {\cal B}_{n_0}} are regular.Proposition REF can now be applied to this special choice of regular functions:f_{\mathbf {n}} &= _{n_0}(1) + \frac{1}{s_{...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.053343117237091064, 0.02401660941541195, -0.01278647966682911, -0.01939333602786064, -0.006809029262512922, 0.012343987822532654, 0.017058812081813812, 0.007213375065475702, 0.015555866062641144, 0.034819506108760834, -0.011611588299274445, 0.022856976836919785, -0.026961466297507286, 0...
a714e7882f1c2fdbf5b3488d821d87c0fc49e8a3
subsection
68
160
Lower and upper matching bounds
Define on [0,+\infty [ \times [0,\rho _1] the function \psi (r_1, q_1)f(r_1) + g(q_1) - \frac{\alpha _1}{2} r_1 q_1 where f: [0,+\infty [ \rightarrow \mathbb {R} and g: [0,\rho _1] \rightarrow \mathbb {R} are such thatf(r_1) \!\!{\sup }_{q_0 \in [0,\rho _0]} {\inf }_{r_0 \ge 0} \Big \lbrace \psi _{P_0}(r_0) + \alpha _1...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.016009269282221794, 0.027653761208057404, -0.009729179553687572, -0.007726113311946392, 0.011522401124238968, 0.012628856115043163, 0.018252704292535782, 0.04090070724487305, 0.015994008630514145, 0.04209110140800476, -0.008531155064702034, -0.012018398381769657, -0.03610860928893089, 0...
2315b666ca61b2ba881d2aa70bb8ad2d03d51a10
subsection
69
160
Lower and upper matching bounds
Hence\forall r_1 \ge : \inf _{q_1 \ge [0,\rho _1]} \psi (r_1, q_1) = \psi (r_1, \rho _1) \,.The latter implies that for all r_1 \ge we have\inf _{q_1 \ge [0,\rho _1]} \psi (r_1, q_1) - \inf _{q_1 \ge [0,\rho _1]} \psi (, q_1) &= \psi (r_1, \rho _1) - \psi (, \rho _1)\\ &= f(r_1) - f() - \frac{\alpha _1 \rho _1}{2}(r_1 ...
{ "cite_spans": [] }
10.1088/1742-5468/ab3430
1805.09785
Entropy and mutual information in models of deep neural networks
[ "Marylou Gabrié", "Andre Manoel", "Clément Luneau", "Jean Barbier", "Nicolas Macris", "Florent Krzakala", "Lenka Zdeborová" ]
[ "cs.LG", "cond-mat.dis-nn", "cs.IT", "math.IT", "stat.ML" ]
2,018
en
Computer Science
[ -0.007677751127630472, 0.04625720903277397, 0.009680145420134068, -0.0090317502617836, -0.004508247133344412, 0.007483232766389847, 0.042412612587213516, 0.04756925627589226, 0.014821530319750309, 0.0143180713057518, -0.009954758919775486, -0.02952100895345211, -0.046623360365629196, 0.013...