Computational Theories of the Brain 2018
2018-04-16..2018-04-23
Berkeley, United States
[Symposium on Visions of the Theory of Computing 2013, Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives, The Role of TCS in Modern Machine Learning, Emerging Challenges in Deep Learning (Simons Institute workshop, 2019), PyData Berlin 2017, Meet the Fellows Welcome Event]
as the organisers put it
This workshop is about general computational principles for networks of neurons that help us understand experimental data, about principles that enable us to reproduce aspects of the brain's astounding computational capability in models and neuromorphic hardware, and about the connections between computational neuroscience and machine learning. Support is gratefully acknowledged from:
| title | author | length |
|---|---|---|
| The Prefrontal Cortex as a Meta-Reinforcement Learning System | Matthew Botvinick | 55m |
Claim: prefrontal cortex implements a learned learning algorithm — a meta‑reinforcement learner — realized as activity dynamics rather than ongoing synaptic… | ||
| Predictive Coding Models of Perception | David Cox | 52m |
Argues that training convolutional autoencoders with an LSTM bottleneck to do future-frame prediction yields internal codes that decode identity, motion and… | ||
| The Brain as an Efficient, Robust Adaptive Learner | Sophie Denève | 51m |
Claim: networks of integrate-and-fire neurons learn and implement low‑dimensional dynamical systems using only local signals, by combining a fast inhibitory… | ||
| Towards Biologically Plausible Deep Learning: ... | Asja Fischer | 35m |
Leaky‑integrator neurons are shown to perform Langevin MCMC inference on an energy‑based model whose update direction equals the energy gradient, and the first… | ||
| Emergent elasticity in the neural code for space | Surya Ganguli | 54m |
Evidence shows grid code coherence decays within 40 seconds without landmarks, so a solvable attractor-plus-Hebbian model is built: landmark synapses learn by… | ||
| Does the Neocortex Use Grid Cell-Like Mechanisms to Learn the Structure of Objects? | Jeff Hawkins | 56m |
A concrete claim: the neocortex is a location processor where every cortical column uses grid cell-like location signals (proposed in layer 6A) to learn… | ||
| Dynamic Neural Network Structures Through Stochastic Rewiring | Robert Legenstein | 47m |
Argues that constant, largely stochastic synaptic rewiring implements a sampler over network parameters so circuits sample from an objective distribution… | ||
| Backpropagation and Deep Learning in the Brain | Timothy Lillicrap | 01h 00m |
Argues that brains likely send detailed, neuron-specific feedback for credit assignment and that some nonstandard algorithms can approximate backprop: feedback… | ||
| An Algorithmic Theory of Brain Networks | Nancy Lynch | 49m |
A formal stochastic spiking neural network model uses synchronized discrete rounds, sigmoid noisy thresholds and signed weights to prove concrete algorithmic… | ||
| Networks of Spiking Neurons Learn to Learn and Remember | Wolfgang Dr. Maass (University of St. Gallen) | 55m |
New networks of spiking neurons add an adaptive firing thresholds term (simple decay-plus-spike increment) and use rewiring during training plus… | ||
| Plenary Discussion: What Is Missing in Current Theories of Brain Computation? | 01h 03m | |
The discussion pinpoints concrete gaps: no working algorithm explains how the brain amplifies tiny sensory inputs into rich 3D perception or stores tens of… | ||
| Functional Triplet Motifs Underlie Accurate Predictions of Single-Trial Responses... | Jason MacLean | 47m |
High-speed two-photon targeted line scans sample up to ~350 excitatory layer 2/3 neurons at ~30 Hz and across datasets (~4,500 neurons) to predict single-trial… | ||
| The Sparse Manifold Transform | Bruno Olshausen | 54m |
Proposes a Sparse Manifold Transform: compute L1 sparse codes of local image patches (e.g. 16×16), learn an embedding of the learned dictionary (hundreds to… | ||
| Playing Newton: Automatic Construction of Phenomenological, Data-Driven Theories and Models | Ilya Nemenman | 51m |
Automated inference builds phenomenological, data-driven theories as nested ODE families (sigmoidal-network and S-system bases) and scores them with Bayesian… | ||
| A Functional Classification of Glutamatergic Circuits in Cortex and Thalamus | S. Murray Sherman | 51m |
Using slice physiology with laser glutamate uncaging, the talk claims glutamatergic synapses split into two functional classes: drivers (large, proximal… | ||
| On the Link Between Energy & Information for the Design of Neuromorphic Systems | Narayan Srinivasa | 42m |
Argument links energy and information to build edge neuromorphic hardware: Eta Compute demonstrates an asynchronous, event-driven chip in 55 nm that supports… | ||
| Plenary Discussion: How Can One Test/Falsify Current Theories of Brain Computation? | Nancy Kanwisher | 53m |
Lists specific experiments, datasets and failure conditions: use the CRCNS repository and Allen Brain Observatory to test models, demand mechanistic… | ||
| Control of Synaptic Plasticity in Deep Cortical Networks | Pieter Roelfsema | 48m |
Claim: feedback signals label only the neurons that will change, turning late apical dendrite calcium events into an eligibility trace that, combined with a… | ||
| Computation with Assemblies | Christos Papadimitriou | 55m |
Asserts an Assembly Hypothesis where synchronized, persistent neuron populations perform operations like projection, association and merge; projection creates… | ||
| Capacity of Neural Networks for Lifelong Learning of Composable Tasks | Les Valiant | 47m |
Proposes a concrete neural circuit model using a random graph spike-timing model with parameters N (neurons), D (fan‑out), K (synaptic strength) and R (neurons… | ||
| An Integrated Cognitive Architecture | Greg Wayne | 53m |
Builds agents that combine a content-addressable external memory (originally Differentiable Neural Computer / Neural Turing Machine) with learned latent state… | ||
| Cal Day Lecture: Can a Machine be Conscious? Towards a Conscious AI | Manuel Blum | 59m |
Defines a Conscious Turing Machine whose tiny short-term memory broadcast sends a single small “stage” to ~10^11 unconscious processors; each processor… | ||
| Personalization and Retail: Lessons from Stitch Fix | Brad Klingenberg | 25m |
Stitch Fix runs a two-stage human+machine platform that picks five items per “fix,” charges a $20 fee, and uses thousands of remote employee stylists plus an… | ||
| Frontiers of Machine Learning from the Lens of Retail | Devavrat Shah | 22m |
Company vision treats extreme sparsity as the central technical problem and claims accurate per-SKU, per-channel demand forecasts by stitching text, image and… | ||
| Current and Future Trends in Mapping | Vinay Shet | 25m |
Lyft handles about 10 million rides/week and covers roughly 95% US population, and mapping work focuses on two paths: an open network for partners plus an… | ||
| Designing and Delivering Experiences using Machine Learning and AI | Anil Kamath | 22m |
Adobe describes a SaaS platform called the Experience Cloud that collects high‑dimensional, real‑time customer signals (web, app, ads, call centers), stores… | ||
| The Economics of Bitcoin | Ciamac Moallemi | 28m |
Frames Bitcoin as a two sided platform and builds a queueing game with analytic equilibrium solutions fitted to six years of daily blockchain data, using… | ||
| Making Blockchains more Useful for Business | Dan Boneh | 20m |
Argues that the UTXO model is a poor fit for business because addresses and amounts are public, and proposes hiding values with commitments plus zero‑knowledge… | ||
| Real-Time Decision Making in Financial Networks | Balaji Prabhakar | 26m |
A clock‑synchronization system called Huygens synchronizes thousands of servers to single‑digit nanoseconds using probes and a reconstruction pipeline, has… | ||
| nothing matches | ||