Chronological feed of everything captured from Demis Hassabis.
AlphaZero generalizes AlphaGo Zero's tabula rasa reinforcement learning to a single algorithm achieving superhuman performance in chess, shogi, and Go. Starting from random play with only game rules, it reaches superhuman levels within 24 hours and defeats world-champion programs in each domain. This eliminates reliance on human expertise, search heuristics, or handcrafted evaluations traditional in chess AI.
WaveNet achieves state-of-the-art realistic speech synthesis across languages but is limited by sequential autoregressive generation, unsuitable for real-time parallel computing. Probability Density Distillation trains a parallel feed-forward network from a pretrained WaveNet, preserving audio quality. The distilled model generates speech over 20x faster than real-time and powers Google Assistant's English and Japanese voices.
Imagination-Augmented Agents (I2As) fuse model-free and model-based deep RL by feeding learned environment model predictions as context into policy networks, enabling arbitrary implicit planning without prescribed model usage. This contrasts with traditional model-based methods that rigidly dictate model-to-policy translation. I2As demonstrate superior data efficiency, performance, and robustness to model errors over baselines.
SCAN learns compositional, hierarchical visual concepts through fast symbol association grounded in unsupervised disentangled primitives, requiring minimal symbol-image pairings without assumptions on symbol form. It supports bi-directional multimodal inference, generating diverse images from symbols and symbols from images. Learned logical operations allow traversal, manipulation, and recombination of concepts to produce novel visuals beyond training distributions.
NoisyNet introduces parametric noise to deep RL agent weights, inducing stochastic policies that enhance exploration without relying on traditional heuristics. Noise parameters are optimized via gradient descent alongside network weights, maintaining low computational overhead. It outperforms entropy rewards in A3C and ε-greedy in DQN/dueling architectures across Atari games, often achieving superhuman performance from subhuman baselines.
An AI agent learns to interpret natural language instructions in a simulated 3D environment using reinforcement learning combined with unsupervised learning, starting from minimal priors. It grounds linguistic symbols to emergent perceptual representations and action sequences, enabling execution of unseen instructions in novel situations. Learning efficiency for new words accelerates with growing semantic knowledge, demonstrating semantic bootstrapping.
Neural Episodic Control (NEC) introduces a deep RL agent that rapidly assimilates new experiences using a semi-tabular value function representation: a buffer of slowly changing state embeddings paired with rapidly updated value estimates. This design enables sample efficiency far superior to standard deep RL methods, which require orders of magnitude more data than humans. Empirical results demonstrate NEC learns significantly faster than state-of-the-art general-purpose deep RL agents across diverse environments.
DeepMind Lab is a first-person 3D game platform built for advancing general AI and ML through agent training in large, partially observed, visually diverse worlds. It features a simple, flexible API that supports rapid iteration on creative task designs and novel AI architectures. Powered by a high-performance game engine, it is optimized for research community use.
Neural networks suffer from catastrophic forgetting when learning sequential tasks, overwriting prior knowledge. The proposed method overcomes this by selectively slowing learning rates on weights critical to old tasks using elastic weight consolidation. It demonstrates scalability on permuted MNIST classification sequences and effectiveness on sequential Atari 2600 games, maintaining long-term expertise.
A model-free episodic control algorithm mimicking hippocampal episodic memory enables rapid exploitation of rewarding environment nuances, contrasting with deep RL's millions of interactions for human-level performance. It solves difficult sequential decision tasks faster than state-of-the-art deep RL methods. On challenging domains, it achieves both accelerated strategy acquisition and superior total rewards.
Proposes algorithmic framework modeling perception-memory interface using approximate Hubel-Wiesel modules for hierarchical invariant object representations. Accounts for view-dependence/invariance in visual cortex (V1 to IT) and episodic interference in medial temporal lobe. Unifies two-speed learning from neocortex to hippocampus via novel interpretation of Hubel-Wiesel conjecture on complex receptive fields.