Senior staff research scientist at DeepMind. Opinions are my own. Re-tweets and favorites not to be considered as endorsements.
London, England
- Tweets 9,368
- Following 1,391
- Followers 6,527
- Likes 15,641
Pinned Tweet
Excited to share this podcast - where we got to talk about the current trajectory of AI agent research, and open challenges in enabling safe and reliable multi-agent coordination at scale.
What happens when millions of AI agents start negotiating, transacting, and delegating to one another?
@weballergy joined our podcast with @fryrsquared to explore the rise of agentic economies – and how we can diversify agent decision-making to avoid AI groupthink.
Timecodes:
00:00 Intro
1:07 Defining AI agents
4:44 Agentic exploration in science and research
15:46 Delegation between agents
22:46 Agentic security and traps
29:31 Building an agentic economy
33:22 Cognitive monoculture
36:29 Distributed intelligence
5
27
3,118
Deep learning models are often perceived as black boxes. In our most recent work, Acquisition of Chess Knowledge in AlphaZero arxiv.org/abs/2111.09259 , we try to unpack how AlphaZero represents knowledge, where it resides within the network, and when it is acquired in training
3
92
426
I'm happy to share that I got promoted to the role of Senior Staff Research Scientist here at Google DeepMind.
It's been an incredibly exciting year, though the truly exciting work, as always, lies ahead.
17
2
355
31,357
I'm excited to share a new paper: "Mastering Board Games by External and Internal Planning with Language Models"
storage.googleapis.com/deepm…
(also soon to be up on Arxiv, once it's been processed there)
5
54
344
152,291
'Adversarial Generation of Natural Language': producing realistic sentences arxiv.org/abs/1705.10929 #deeplearning #machinelearning #NLP #AI
132
253
DeepMoji: Predicting emojis for classifying text sentiment/emotion/sarcasm arxiv.org/abs/1708.00524 #NLP #deeplearning #AI
4
125
247
'Relational recurrent neural networks': performing complex relational reasoning in memory networks. arxiv.org/abs/1806.01822 #DeepLearning #AI #MachineLearning
75
224
"To Trust Or Not To Trust A Classifier" by Google Research arxiv.org/abs/1805.11783 : beyond simple confidence scores. The ability to auto-detect bad predictions in critical for safe deployments in sensitive applications. #MachineLearning #DataScience #AI
1
64
214
Proud to share the results of our work on applying deep learning for early prediction of future acute kidney injury from electronic health records in our collaboration with the US Department of Veterans Affairs - just published in Nature: nature.com/articles/s41586-0…
9
47
191
'Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning' arxiv.org/abs/1709.00103 #MachineLearning
2
74
179
'A Probabilistic U-net for Segmentation of Ambiguous Images': a cool new paper by my colleagues at DeepMind on how to deal with uncertainty in segmentation models. arxiv.org/abs/1806.05034 #DeepLearning #MachineLearning
2
68
167
'Hyperbolic Entailment Cones for Learning Hierarchical Embeddings': viewing hierarchical relations as partial orders based on a family of nested
geodesically convex cones arxiv.org/abs/1804.01882 #AI #MachineLearning
1
34
151
'Dilated Convolutions for Modeling Long-Distance Genomic Dependencies' arxiv.org/abs/1710.01278 #DeepLearning #Genomics #AI
2
43
144
'A Distributional Perspective on Reinforcement Learning': modeling the full distribution of return. arxiv.org/abs/1707.06887 #machinelearning
58
143
Happy to announce our new paper "Advancing mathematics by guiding human intuition with AI" nature.com/articles/s41586-0… that was just published in Nature, and one of the associated maths papers "The signature and cusp geometry of hyperbolic knots" arxiv.org/abs/2111.15323
4
24
126
'Twin Networks: Using the Future as a Regularizer': arxiv.org/abs/1708.06742 #DeepLearning #MachineLearning #AI
2
47
126
'MDNet: A Semantically and Visually Interpretable Medical Image Diagnosis Network' arxiv.org/abs/1707.02485 #deeplearning #AI #medicine
1
57
131
'Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks' arxiv.org/abs/1707.01836 #deeplearning #AI #machinelearning
2
57
118
'Super Convergence: Very Fast Training of Residual Networks Using Large Learning Rates' arxiv.org/abs/1708.07120 #deeplearning #AI
41
111
'Character-level Recurrent Neural Networks in Practice: Comparing Training and Sampling Schemes' arxiv.org/abs/1801.00632 #DeepLearning #MachineLearning #AI
28
108
'Boundary-Seeking Generative Adversarial Networks': Nice new results by Yoshua Bengio's group. arxiv.org/abs/1702.08431 #deeplearning #AI
1
53
111
'Deep Learning is Robust to Massive Label Noise': encouraging results. arxiv.org/abs/1705.10694 #deeplearning #machinelearning #AI
45
101
'The Consciousness Prior': a short write-up by Yoshua Bengio arxiv.org/abs/1709.08568 Thought-provoking, but no experiments given. #AI
2
33
97
'Unsupervised Representation Learning by Sorting Sequences': learning promising visual representations arxiv.org/abs/1708.01246 #deeplearning
26
97
Fairness for Unobserved Characteristics: Insights from Technological Impacts on Queer Communities arxiv.org/abs/2102.04257 : highlighting some of the key challenges in ensuring algorithmic fairness for queer communities @jackayline @empiricallykev @Shakir_za
2
28
84
'A systematic study of the class imbalance problem in convolutional neural networks': advocating for oversampling arxiv.org/abs/1710.05381
2
38
89
'Sparsely Connected Convolutional Networks' arxiv.org/abs/1801.05895 : Introducing SparseNets, a child of DenseNets and ResNets. The authors claim that the method can outperform these with fewer parameters. #DeepLearning #MachineLearning #ComputerVision #AI
29
89
'Dilated Recurrent Neural Networks': dilated recurrent skip connections. arxiv.org/abs/1710.02224 #DeepLearning #MachineLearning #AI
1
20
84
'Visualizing LSTM Decisions': on interpretability in recurrent neural networks. arxiv.org/abs/1705.08153 #deeplearning #machinelearning #AI
37
82
'Stochastic Gradient Descent as Approximate Bayesian Inference': arxiv.org/abs/1704.04289 #machinelearning
1
29
82
'Deep Episodic Memory: Encoding, Recalling, and Predicting Episodic Experiences for Robot Action Execution' arxiv.org/abs/1801.04134 #DeepLearning #Robotics #AI
22
82
'Grounded Language Learning in a Simulated 3D World': rewards for following written instructions. arxiv.org/abs/1706.06551 #AI #deeplearning
23
80
'Billion-scale similarity search with GPUs': kNN search that is 8.5x faster than prior state-of-the-art. (by FAIR) arxiv.org/abs/1702.08734
29
77
'Noisy Networks for Exploration': on learning parametric noise for exploration in RL agents. arxiv.org/abs/1706.10295 #deeplearning #AI
19
76
Happy to announce our upcoming NeurIPS 2022 workshop on "A Participatory Approach to AI for Mental Health" sites.google.com/view/pai4mh… , hoping to bring experts and communities together to jointly shape the vision for how tech can help with wellbeing and mental health
4
21
73
'Non-local Neural Networks' by CMU, FAIR arxiv.org/abs/1711.07971 #DeepLearning #MachineLearning #AI
32
76
'The Devil is in the Decoder': the choice of decoder matters in pixel-wise prediction tasks arxiv.org/abs/1707.05847 #deeplearning #AI
24
72
'Swish: a Self-Gated Activation Function': a replacement for RELU-s? By Google Brain. arxiv.org/abs/1710.05941 #DeepLearning #MachineLearning
48
75
'Learning to Search with MCTSnets' arxiv.org/abs/1802.04697 : learning where, what and how to search. #MachineLearning #AI
1
12
71
I am happy to advertise a position for a student researcher under the Google DeepMind program deepmind.google/about/career… , for a project with @TZahavy , myself, and others, at the intersection of generative modelling, creativity, and planning. Strong programming skills required.
2
12
70
14,437
'Programmable Agents': a new paper by colleagues from DeepMind arxiv.org/abs/1706.06383 #deeplearning #AI
2
31
69
'Auto-Differentiating Linear Algebra' arxiv.org/abs/1710.08717 #MachineLearning
1
25
69
'Adversarial Spheres' by Google Brain: towards a better understanding of adversarial examples in #DeepLearning arxiv.org/abs/1801.02774 "the vulnerability of neural networks to small adversarial perturbations is a logical
consequence of the amount of test error observed" #AI
1
14
67
Applied AI solutions require setting up lasting interdisciplinary partnerships with domain experts, and here we share some thoughts and guidelines on forming these collaborations.
"AI for social good: unlocking the opportunity for positive impact" nature.com/articles/s41467-0…
2
16
61
'First-order Methods Almost Always Avoid Saddle Points' arxiv.org/abs/1710.07406 Can't argue with gradient descent. #DeepLearning #AI
20
57
'Factorization tricks for LSTM networks': speeds ups and fewer parameters. arxiv.org/abs/1703.10722 #deeplearning #machinelearning
25
60
'Adversarial Transformation Networks: Learning to Generate Adversarial Examples' by Google Research arxiv.org/abs/1703.09387 #deeplearning
17
58
'A Structured Self-attentive Sentence Embedding' (by Yoshua Bengio et al.): 2D matrix embedding + self-attention arxiv.org/abs/1703.03130 #AI
1
12
56
'Panoptic Segmentation': proposing a new hybrid between instance and semantic segmentation tasks. arxiv.org/abs/1801.00868 #DeepLearning #ComputerVision #MachineLearning #AI
1
15
54
'Do Convolutional Neural Networks Act as Compositional Nearest Neighbors?'; Experiments on semantic segmentation and image-to-image translation. arxiv.org/abs/1711.10683 #MachineLearning #DeepLearning #AI
23
53
'Polar Transformer Networks': invariance to translation, and equivariance to rotation and scale. arxiv.org/abs/1709.01889 #DeepLearning #AI
1
21
56
Today we released a detailed protocol of the process that we used to develop acute kidney injury (AKI) risk prediction models: "Use of deep learning to develop continuous-risk models for adverse event prediction from electronic health records" rdcu.be/cj1vf
1
14
52
'Opportunities in Machine Learning for Healthcare' arxiv.org/abs/1806.00388 : many open challenges and a huge potential for impact. #MachineLearning #Health #AI
15
56
'Contextual Explanation Networks': interpretability as regularization. arxiv.org/abs/1705.10301 #deeplearning #machinelearning #AI
22
47
'Neural Embeddings of Graphs in Hyperbolic Space' arxiv.org/abs/1705.10359 : worth comparing with: arxiv.org/abs/1705.08039 - similar ideas.
18
50
'Causal Generative Neural Networks' arxiv.org/abs/1711.08936 : 'Unlike previous approaches, the generative networks used in CGNN allow non-additive noise terms to model flexible conditional distributions' #DeepLearning #MachineLearning #AI
27
49
Really enjoying the AI+HCI workshop at @icmlconf - great talks and important topics to reflect on as a field. We need to be thinking much more about complex interactions and wider unintended consequences of AI deployment, and invest more in sociotechnical research.
1
11
46
12,259
'Recurrent Ladder Networks': a recurrent extension of the Ladder network arxiv.org/abs/1707.09219 #deeplearning #machinelearning #AI
1
24
49
'Beyond Finite Layer Neural Networks: Bridging Deep Architectures and Numerical Differential Equations' arxiv.org/abs/1710.10121 #AI
1
16
46
'Neural Block Sampling': a neural approach to automate MC proposal
construction. arxiv.org/abs/1708.06040 #machinelearning #statistics
20
45
Happy to share that our paper "Detecting shortcut learning for fair medical AI using shortcut testing" has just been published in Nature Communications. We validate our method on clinical ML tasks in radiology and dermatology. nature.com/articles/s41467-0…
1
12
43
4,792
'Variance Regularizing Adversarial Learning': an interesting read. arxiv.org/abs/1707.00309 #deeplearning #machinelearning #AI
20
44
'Time Series Segmentation through Automatic Feature Learning' arxiv.org/abs/1801.05394 : Using #DeepLearning to detect abrupt changes in trends in time series data. #MachineLearning #AI
13
43
Proud to see our first piece of work at Health Research here at DeepMind appear at Nature Medicine. It's been a long journey and a pleasure to work with quite an inspiring team. #AI #DeepLearning #Healthcare
Teams at @DeepMind_Health and @Moorfields have developed AI technology that can detect eye disease and prioritise patients. 'Clinically applicable deep learning for diagnosis and referral in retinal OCT' has been published online in @NatureMedicine today: dx.doi.org/10.1038/s41591-01…
1
8
42
GANs + reinforcement learning = OR-GAN; A new paper from Harvard. arxiv.org/abs/1705.10843 #deeplearning #machinelearning #AI
1
24
43
'Imagination-Augmented Agents for Deep Reinforcement Learning' by colleagues at DeepMind: arxiv.org/abs/1707.06203 #deeplearning #AI
11
44
Backprop Evolution arxiv.org/abs/1808.02822 : evolutionary approach towards finding alternative update equations to potentially supplant the standard backprop update. #AI #MachineLearning #DeepLearning
11
44
As a community, we need to recognize the value of using AI systems not only as tools for 'solving problems' and automating processes - but rather, models about the world from which we ourselves can learn and improve - having them augment our abilities rather than displace them
4
42
'Tactics of Adversarial Attack on Deep Reinforcement Learning Agents' arxiv.org/abs/1703.06748 #deeplearning #machinelearning #AI
18
42
'Deep Reinforcement Learning from Human Preferences' by DeepMind and OpenAI arxiv.org/abs/1706.03741 #DeepLearning #ReinforcementLearning #AI
11
42
'Predicting cancer outcomes from histology and genomics using convolutional networks' biorxiv.org/content/early/20… #DeepLearning #Medicine
17
38
'Hidden Physics Models: Machine Learning of Nonlinear Partial Differential Equations' arxiv.org/abs/1708.00588 #physics #machinelearning
24
39
I'm really excited to share the work on Med-Gemini, which brings and unites a number of advances in LLM reasoning, search integration, long context utilization, and multimodal understanding into the medical domain, unlocking new opportunities.
Introducing Med-Gemini, a family of models that extends the best of Gemini into medicine! ✨⚕️
Highlights of what you can do with Med-Gemini:
> Answer medical questions with up-to-date knowledge using agentic web search 🔎❤️🩹
> Converse about your medical images, videos, and long multi-visit health records 📷📹📃
> Do a literature search by uploading tens of biomedical papers and asking questions 📚
> And so much more!
🏗️ Development of Med-Gemini included:
> Advancing clinical reasoning with self-training and search
> Improving multimodal understanding with fine-tuning
> Leveraging long-context capabilities with chain-of-reasoning
Paper: arxiv.org/abs/2404.18416
Below ⬇️, I talk more about self-training with web search to improve Gemini’s clinical reasoning.
1
6
39
4,221
'Hierarchical Disentangled Representations': on leaning independent factors of variation arxiv.org/abs/1804.02086 #DeepLearning #MachineLearning #AI
5
41
'Are GANs Created Equal? A Large-Scale Study' arxiv.org/abs/1711.10337 The study does not find any GAN that consistently outperforms the others. Improvements mostly due to better hyperparam tuning. #DeepLearning #MachineLearning #AI
24
41
"Can Computers Create Art?" arxiv.org/abs/1801.04486 : on the interplay between #technology and #art and future prospects for building genuinely creative #AI
4
19
37
'Training RNNs as Fast as CNNs': SRUs enable fast parallel computations. arxiv.org/abs/1709.02755 #DeepLearning #MachineLearning #AI
11
38
"The Blessings of Multiple Causes": performing causal inference in multiple-cause settings. Inferring latent variables for unobserved confounders. arxiv.org/abs/1805.06826 #MachineLearning #DataScience
11
39
"How do Humans Understand Explanations from Machine Learning Systems? An Evaluation of the Human-Interpretability of Explanation": which properties of explanations are most useful to people? arxiv.org/abs/1802.00682 #MachineLearning #AI
16
38
'Toward Fairness in AI for People with Disabilities: A Research Roadmap' arxiv.org/abs/1907.02227 highlights the potential of #AI in improving the lives of people with disabilities, while noting that many existing AI systems may not be designed appropriately to achieve this goal.
2
13
34
'MINE: Mutual Information Neural Estimation': Introduces MINE-GAN. Has better coverage, fast convergence and fewer mode collapses. arxiv.org/abs/1801.04062 #DeepLearning #MachineLearning #AI
10
38
A nice post on generating high-level document summaries with RNN-s, discussing Pointer-Generator Networks #nlp #deeplearning #AI
Better neural abstractive summarization—@abigail_e_see. Great blog post abigailsee.com/2017/04/16/ta…—Final @acl2017 paper arxiv.org/abs/1704.04368
1
24
35
'Pixel Deconvolutional Networks': improving semantic segmentation. arxiv.org/abs/1705.06820 #deeplearning #machinelearning #computervision
1
23
36
'Generating and designing DNA with deep generative models': using #DeepLearning for designing probes for protein binding microarrays. arxiv.org/abs/1712.06148 #AI #Genomics
14
34
'Learning Awareness Models': building representations of external objects based only on the internal body signals arxiv.org/abs/1804.06318 #AI #MachineLearning #DeepLearning by CMU, UoM, DeepMind, CIFAR
8
34
'Fraternal Dropout' arxiv.org/abs/1711.00066 - an interesting concept. Two RNN copies with different dropout masks forced to be similar.
9
33
Non-differentiable, content-addressable memory: 'A Growing Long-term Episodic & Semantic Memory' arxiv.org/abs/1610.06402 #deeplearning #AI
1
21
35
Working towards improving ML explainability in medical applications is key to ensure that the models are not relying merely on spurious correlations - and high-level conceptual expanations can be hepful for clinical interrogation of model behavior.
Happy to share our work on demystifying recurrent neural nets for medical applications: Concept-based model explanations for Electronic Health Records arxiv.org/abs/2012.02308
@d_mincu, @shaobohou, @martin_sen, @weballergy, @alan_karthi & @JessicaSchrouff #CHIL2021
9
34
Definitely one of the most fun and interesting projects I've had the opportunity to work on! Working with Vladimir Kramnik had been a privilege - seeing how we can use AlphaZero for experimenting with alterations to the rules of chess, exploring some exciting new chess variants.
In a bid to explore new frontiers in chess, our researchers worked with Vladimir Kramnik to use AlphaZero to test nine new variants of chess.
The result? A more creative and collaborative relationship between chess players and machines.
bit.ly/32fsmxB via @Wired
1
7
34
'Teaching Machines to Describe Images via Natural Language Feedback' arxiv.org/abs/1706.00130 #deeplearning #machinelearning #AI
19
32
"A Deep Reinforcement Learning Chatbot (Short Version)": Introducing MILABOT, a chatbot developed for the Amazon Alexa Prize competition. arxiv.org/abs/1801.06700 #DeepLearning #NLP #AI #MachineLearning
16
34
'Independently Controllable Features': learning disentangled representations by interacting with the environment arxiv.org/abs/1708.01289 #AI
13
32
'Machine learning & artificial intelligence in the quantum domain': an overview of recent advances. arxiv.org/abs/1709.02779 #AI #physics
19
30
Excited to share our latest piece of work in collaboration with Liverpool FC, on developing a football AI assistant - TacticAI, that can help understand and improve corner kick tactics.
We're announcing TacticAI: an AI assistant capable of offering insights to football experts on corner kicks. ⚽
Developed with @LFC, it can help teams sample alternative player setups to evaluate possible outcomes, and achieves state-of-the-art results. 🧵 dpmd.ai/49PGq1b
1
1
31
3,569
'A Capacity Scaling Law for Artificial Neural Networks': computing the VC and the MacKay dimension arxiv.org/abs/1708.06019 #deeplearning #AI
7
28
If you are as interested as we are in exploring how planning and reasoning with large language models can help master board games and vice versa,
Join us this coming Wednesday in Vancouver at the GDM NeurIPS booth, where we will be showing a demo.
deepmind.google/discover/eve…
I'm excited to share a new paper: "Mastering Board Games by External and Internal Planning with Language Models"
storage.googleapis.com/deepm…
(also soon to be up on Arxiv, once it's been processed there)
6
30
5,061
Fader Networks (by FAIR): disentangling image attributes and adjusting their strength arxiv.org/abs/1706.00409 #deeplearning
1
18
28
A nice write-up on an important paper. On how everything can be viewed as a model and how all/most data products (and data structures) can and should be optimized based on the statistical properties of the data. An ML-first dev world.
New blog post: On the lessons from the "Learned Index" paper and its impact on product thinking/building. deliprao.com/archives/262
1
6
26
'Multiscale sequence modeling with a learned dictionary' arxiv.org/abs/1707.00762 #deeplearning #machinelearning #AI
7
26
On how we recognize faces. (Un)surprisingly similar to modern #AI approaches / encodings.
TL;DR Primate brains use what amounts to "face2vec" to encode facial identity + it's decodable from neural activity!
cell.com/cell/fulltext/S0092…
1
17
29
'Generator Reversal': modeling natural code distributions in deep generative models. arxiv.org/abs/1707.09241 #deeplearning #machinelearning
1
12
28