BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Microsoft Cambridge AI Pizza Event on 25th April - Speaker to be c
 onfirmed
DTSTART:20240425T163000Z
DTEND:20240425T180000Z
UID:TALK216226@talks.cam.ac.uk
CONTACT:Kimberly Cole
DESCRIPTION:Dear Cambridge AI & Machine Learning Enthusiast\,\n \nWe're th
 rilled to announce our next eagerly-anticipated AI & Pizza event! Mark you
 r calendars for an insightful evening at 5:30 pm on Thursday\, April 25th\
 , 2024. It's your chance to savour the latest advancements in AI and machi
 ne learning\, right here in Cambridge. Of course\, Pizza will be provided 
 after the talk!\n \nLocation: The Auditorium\, 21 Station Rd\nAgenda:\n5:3
 0 pm – 6:00 pm: Talks\n6:00 pm – 7:00 pm: Networking\, Pizza\, and Ref
 reshments\n \nTalk 1\nTime: 5:30 pm-5:45 pm\nSpeaker: Samuel Holt\nTitle: 
 Extending Large Language Models for Large Code Base Generation and Machine
  Learning Advances in Treatment Effect Analysis\, Continuous-time Control 
 and Symbolic Regression.\nAbstract:\nThis talk covers (1) how large langua
 ge models (LLMs) can be extended for large code base generation\, (2) a no
 vel approach to inferring unbiased treatment effects in longitudinal setti
 ngs using a closed-form ordinary differential equation (ODE) instead of tr
 aditional neural network models\, (3) machine learning (ML) advances in mu
 lti-modal transformers that can encode a dataset for symbolic regression\,
  and ML advances in continuous-time control. First (1)\, Transformer-based
  large language models (LLMs) are constrained by the fixed context window 
 of the underlying transformer architecture\, hindering their ability to pr
 oduce long and coherent outputs. Memory-augmented LLMs are a promising sol
 ution\, but current approaches cannot handle long output generation tasks 
 since they (1) only focus on reading memory and reduce its evolution to th
 e concatenation of new memories or (2) use very specialized memories that 
 cannot adapt to other domains. This paper presents L2MAC\, the first pract
 ical LLM-based stored-program automatic computer (von Neumann architecture
 ) framework\, an LLM-based multi-agent system\, for long and consistent ou
 tput generation. Its memory has two components: the instruction registry\,
  which is populated with a prompt program to solve the user-given task\, a
 nd a file store\, which will contain the final and intermediate outputs. E
 ach instruction in turn is executed by a separate LLM agent\, whose contex
 t is managed by a control unit capable of precise memory reading and writi
 ng to ensure effective interaction with the file store. These components e
 nable L2MAC to generate extensive outputs\, bypassing the constraints of t
 he finite context window while producing outputs that fulfill a complex us
 er-specified task. We empirically demonstrate that L2MAC achieves state-of
 -the-art performance in generating large codebases for system design tasks
 \, including HumanEval\, significantly outperforming other coding methods 
 in implementing the detailed user-specified task\, and we provide valuable
  insights into the reasons for this performance gap. Second (2)\, inferrin
 g unbiased treatment effects has received widespread attention in the mach
 ine learning community. In recent years\, our community has proposed numer
 ous solutions in standard settings\, high-dimensional treatment settings\,
  and even longitudinal settings. While very diverse\, the solution has mos
 tly relied on neural networks for inference and simultaneous correction of
  assignment bias. New approaches typically build on top of previous approa
 ches by proposing new (or refined) architectures and learning algorithms. 
 However\, the end result -- a neural-network-based inference machine -- re
 mains unchallenged. In this paper\, we introduce a different type of solut
 ion in the longitudinal setting: a closed-form ordinary differential equat
 ion (ODE). While we still rely on continuous optimization to learn an ODE\
 , the resulting inference machine is no longer a neural network. Doing so 
 yields several advantages such as interpretability\, irregular sampling\, 
 and a different set of identification assumptions. Above all\, we consider
  the introduction of a completely new type of solution to be our most impo
 rtant contribution as it may spark entirely new innovations in treatment e
 ffects in general. We facilitate this by formulating our contribution as a
  framework that can transform any ODE discovery method into a treatment ef
 fects method. Third (3)\, we propose a novel multi-modal transformer archi
 tecture that can encode an entire dataset\, trained with PPO and fine-tune
 d at inference time to achieve a new state-of-the-art for the problem of s
 ymbolic regression\, presenting a framework called Deep Generative Symboli
 c Regression. We also propose new state-of-the-art methods for (A) continu
 ous-time control with observation costs and (B) continuous-time control wi
 th fixed delays. Both methods are model-based RL frameworks\, with (A) usi
 ng a probabilistic ensemble dynamics model and (B) using a newly proposed 
 Neural Laplace dynamics model. In summary\, all these works lay exciting f
 oundations for future research in these areas.\n \nTalk 2\nTime: 5:45 pm-6
 :00 pm\nSpeaker: Ira J. S. Shokar\n\nTitle: Extending Deep Learning Emulat
 ion Across Parameter Regimes for Turbulent Flows\n\nAbstract:\n\nGiven the
  computational expense associated with simultaneous multi-task learning\, 
 we leverage fine-tuning to generalise a transformer-based network emulatin
 g dynamical systems across a range of parameters\, rather than ab initio t
 raining for each new parameter. This allows for rapid adaptation of the de
 ep learning model\, that can be used subsequently across a large range of 
 the parameter space or tailored to a specific regime of study. We demonstr
 ate the model's ability to capture the relevant behaviour\, even at parame
 ter values not seen during training. Applied to an idealised model of atmo
 spheric turbulence\, the speed-up provided by the deep learning model over
  numerical integration makes statistical study of rare events in the physi
 cal system computationally feasible.\n\n \nBest regards\,\nWenbo Gong
LOCATION:The Auditorium\, 21 Station Rd
END:VEVENT
END:VCALENDAR