Machine Learning Street Talk (MLST)16 Jun 2022

#77 - Vitaliy Chiley (Cerebras)

Vitaliy Chiley is a Machine Learning Research Engineer at the next-generation computing hardware company Cerebras Systems. We spoke about how DL workloads including sparse workloads can run faster on Cerebras hardware.

[00:00:00] Housekeeping

[00:01:08] Preamble

[00:01:50] Vitaliy Chiley Introduction

[00:03:11] Cerebrus architecture

[00:08:12] Memory management and FLOP utilisation

[00:18:01] Centralised vs decentralised compute architecture

[00:21:12] Sparsity

[00:23:47] Does Sparse NN imply Heterogeneous compute?

[00:29:21] Cost of distributed memory stores?

[00:31:01] Activation vs weight sparsity

[00:37:52] What constitutes a dead weight to be pruned?

[00:39:02] Is it still a saving if we have to choose between weight and activation sparsity?

[00:41:02] Cerebras is a cool place to work

[00:44:05] What is sparsity? Why do we need to start dense?

[00:46:36] Evolutionary algorithms on Cerebras?

[00:47:57] How can we start sparse? Google RIGL

[00:51:44] Inductive priors, why do we need them if we can start sparse?

[00:56:02] Why anthropomorphise inductive priors?

[01:02:13] Could Cerebras run a cyclic computational graph?

[01:03:16] Are NNs locality sensitive hashing tables?

References;

Rigging the Lottery: Making All Tickets Winners [RIGL]

https://arxiv.org/pdf/1911.11134.pdf

[D] DanNet, the CUDA CNN of Dan Ciresan in Jurgen Schmidhuber's team, won 4 image recognition challenges prior to AlexNet

https://www.reddit.com/r/MachineLearning/comments/dwnuwh/d_dannet_the_cuda_cnn_of_dan_ciresan_in_jurgen/

A Spline Theory of Deep Learning [Balestriero]

https://proceedings.mlr.press/v80/balestriero18b.html

Denne episoden er hentet fra en åpen RSS-feed og er ikke publisert av Podme. Den kan derfor inneholde annonser.

Episoder(271)

What Most People Get Wrong About Evolution | Akarsh Kumar

A lot of people in AI treat evolution as a dumb fallback, basically random search for when you can't take a gradient. Akarsh Kumar thinks that is wrong. Selection hangs on to partial solutions, so mut...

10 Okt 47min

How a Voice Agent Learns the Rhythm of Conversation — Shawn Wen

Tsung-Hsien (Shawn) Wen, CTO of PolyAI, tells Tim Scarfe why voice agents are harder than text agents. Voice adds time, and a good conversation depends on adapting to the person on the line, not just ...

1 Okt 1h 10min

Who Checks a Proof No Human Can Read? — Leo de Moura

Leonardo de Moura created Lean and co-created Z3. ---This episode is sponsored by Parallel.Parallel, where agents find answers: web search, extraction and deep research APIs built for AI agents.Start ...

30 Sep 1h 14min

When AI Research Starts Moving Faster Than Human Research - Zhengyao Jiang

Weco let an AI coding agent rewrite the harness around another agent for eight days: its code, prompts and tools, while the underlying language model stayed fixed. Tim Scarfe asks Weco co-founder Zhen...

26 Sep 43min

How Deep Learning Finally Cracked Messy Tables - Frank Hutter

Frank Hutter, co-founder of Prior Labs, talks about TabPFN, a tabular foundation model that makes predictions in a single forward pass, and the research behind it.TabPFN is pre-trained on synthetic da...

23 Sep 1h 53min

Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick

Alexander Mattick is a researcher at Fraunhofer IIS and a PhD researcher at the University of Technology Nuremberg (UTN), and a regular on Yannic Kilcher's Discord. He first came on MLST in 2022, afte...

21 Sep 2h 14min

How Physical AI Learns Across Language, Video and Action — Ming-Yu Liu

The car making a left turn at the start of this episode was never filmed. Cosmos 3 generated it. Ming-Yu Liu, who leads the Cosmos research at NVIDIA, explains how one model can describe a video, gene...

15 Sep 25min

Speech Recognition Is Not a Solved Problem — Pavan Kumar Reddy

Pavan Kumar Reddy leads audio research at Mistral AI. He joins Tim Scarfe for a deep technical tour of Voxtral — and explains why the frontier of deployed voice is still a cascade of specialised model...

14 Sep 1h 42min