Propose open research directions in systems, architectures, and security that can address these challenges and help unlock AI’s potential to improve lives and society
Next generation of AI systems promises to accelerate developments and impact our lives via making decisions on our behalf, often in highly personalized contexts
Upcoming challenges will be exacerbated by the end of the Moore’s Law, which will constrain
the amount of data these technologies can store and process
Details
AI's Recent Success
Massive Amount of Data (Big Data ~ mobile devices, application logs)
Scalable Computer and Software Systems (Big Systems ~ Hadoop, MapReduce, GFS)
Broad Accessibility of these Technologies (OpenSource ~ TensorFlow, MXNet, PyTorch)
Trends
Mission-Critical AI : adapt and learn new skills continually
Personalized AI : learning the preference of individual users while protecting their privacy
AI across Organization : break down the silos of data within one community
AI demands outpacing the Moore's Law : DRAMs, CPU improvements are at the end of Moore's Law, whereas the explosion of data and AI demands is increasing exponentially
Research Opportunities
R1 : Continual Learning
the need for an agent to learn unpredictable events is to have the ability to simulate and predict the outcome before actually taking an action via Simulated Reality.
Deep RL and fully parallelizable structure is needed
R3 : Explainable Decisions
explainable means that one can identify the properties of the input to the AI algorithm that are responsible for the particular output, and can answer counterfactual or “what-if” questions
For example, one may wish to know what features of a particular organ in an X-ray (e.g., size, color, position, form) led to a particular diagnosis and how the diagnosis would change under minor perturbations of those features.
Build AI systems that can support interactive diagnostic analysis, that faithfully replay past executions, and that can help to determine the features of the input that are responsible for a particular decision, possibly by replaying the decision task against past perturbed inputs. More generally, provide systems support for causal inference.
R7 : Domain-specific Hardware
Google's TPU has 15x ~ 30x faster inference than CPU/GPU, and performance per watt is 30x ~ 80x better
R9 : Cloud-edge Systems
Design cloud-edge AI systems that (1) leverage the edge to reduce latency, improve safety and security, and implement intelligent data retention techniques, and (2) leverage the cloud to share data and models across edge devices, train sophisticated computation-intensive models, and take high quality decisions.
Personal Thoughts
My interest is in 4 research topics listed above, out of 9 introduced in the paper
R1 Continual Learning is the ever-lasting goal of personalized AI systems
R3 Explainable Decisions interests me the most. How can we make the system explainable?
R7, R9 Domain-specific Hardware, Cloud-edge Systems are speed related challenges which can push the boundary of latency, throughput to another level
Good analysis of Trends and Challenges in ML systems which naturally lead to research ideas
Abstract
Details
AI's Recent Success
Trends
simulate and predict the outcome
before actually taking an action viaSimulated Reality
.explainable
means that one can identify the properties of the input to the AI algorithm that are responsible for the particular output, and can answer counterfactual or “what-if” questionsBuild AI systems that can support interactive diagnostic analysis, that faithfully replay past executions, and that can help to determine the features of the input that are responsible for a particular decision, possibly by replaying the decision task against past perturbed inputs. More generally, provide systems support for causal inference.
Personal Thoughts
Continual Learning
is the ever-lasting goal of personalized AI systemsExplainable Decisions
interests me the most. How can we make the system explainable?Domain-specific Hardware, Cloud-edge Systems
are speed related challenges which can push the boundary of latency, throughput to another levelLink : https://arxiv.org/pdf/1712.05855v1.pdf Authors : Stoica, et al. 2017