Subramanian "Subbu" Iyer
Sunnyvale, California, United States
7K followers
500+ connections
Articles by Subramanian "Subbu"
-
Attention! Translating transformers to plain English.
Attention! Translating transformers to plain English.
By Subramanian "Subbu" Iyer
-
Embeddings explained in plain English
Embeddings explained in plain English
By Subramanian "Subbu" Iyer
Activity
-
Large language models #LLMs exhibit various skills to different degrees and this has enthralled the whole world leading to the flowering of a million…
Large language models #LLMs exhibit various skills to different degrees and this has enthralled the whole world leading to the flowering of a million…
Shared by Subramanian "Subbu" Iyer
-
Now to the third post that concludes this series "on transitions". This is an unpublished note from my archives, it's from the days when I was a…
Now to the third post that concludes this series "on transitions". This is an unpublished note from my archives, it's from the days when I was a…
Posted by Subramanian "Subbu" Iyer
-
This is the second of three notes "on transitions" Let me excerpt the farewell note I sent to the #dataScience teams (previously part of #AI, before…
This is the second of three notes "on transitions" Let me excerpt the farewell note I sent to the #dataScience teams (previously part of #AI, before…
Posted by Subramanian "Subbu" Iyer
Experience
Education
-
Harvard Business School
The GMP is a comprehensive 4-month full time program covering all aspects of general management.
-
Dissertation: Efficient and Effective Symbolic Model Checking
Advisor: Prof. E Allen Emerson (Turing award 2007) -
All India Rank 32 at the IIT Joint Entrance Exam 1995.
-
Graduate level Mathematics by invitation only for 15 qualifiers of the Indian National Math Olympiad.
Publications
-
The Demand Forecasting Project at Target: Improving Collaboration and Adoption
Foresight: The International Journal of Applied Forecasting
This article provides an update on the status of the Demand Forecasting Engine at Target, a system capable of efficiently generating more than one billion weekly forecasts required by different operations within the company. It describes the challenges faced in securing adoption of its forecasts and improving collaboration among data science, product management, and forecast user teams.
-
Efficient and effective symbolic model checking
The University of Texas at Austin
In this dissertation, we propose techniques to increase the capacity of automatic state-based verification as applied to sequential designs, i.e., symbolic model checking. Firstly, we propose the use of dynamically partitioned ordered Binary Decision Diagrams as a capable data structure. This leads to vast improvements in state space traversal
in general and error detection in buggy designs, in particular. Secondly, we propose a partitioned approach to model checking, which splits the…In this dissertation, we propose techniques to increase the capacity of automatic state-based verification as applied to sequential designs, i.e., symbolic model checking. Firstly, we propose the use of dynamically partitioned ordered Binary Decision Diagrams as a capable data structure. This leads to vast improvements in state space traversal
in general and error detection in buggy designs, in particular. Secondly, we propose a partitioned approach to model checking, which splits the problem into multiple partitions handled independently of each other.
We present a model checking algorithm that aggregates expensive cross-over images by localizing computation to individual partitions. This algorithm is more suited to parallelization than existing model checking approaches. It reduces the number of cross-over images and drastically outperforms extant approaches in terms of cross-over image computation cost as well as total model checking time, often by
two orders of magnitude.
We address the issue of time scalability in verification, whereby the availability of larger amounts of computation time enables greater exploration of the state space. From a practical standpoint, we observe that extant verification approaches are unable to proceed very deep into the state space. It is our conjecture that
partitioning can help in this context and we explore this issue further. Finally, we study the combination of partitioned binary decision diagrams with bounded model checking for more scalable and efficient model checking. We give a technique to scale formal verification to a large grid of processors that demonstrates marked superiority over existing approaches.
The contributions of this dissertation are in improving the capacity of symbolic model checking approaches to formal verification, in terms of time and memory requirements, as well as in the development of techniques that are more readily amenable to parallel and distributed model checking.
Patents
-
Validating one or more circuits using one or more grids
Issued 8181132
In one embodiment, a method includes simulating by one or more computer systems a larger circuit to assign one or more values to one or more latch variables associated with the larger circuit, generating by the one or more computer systems one or more reduced circuits from the larger circuit according to the values assigned to the latch variables, generating by the one or more computer systems a transition relation (TR) for each reduced circuit, and generating by the one or more computer…
In one embodiment, a method includes simulating by one or more computer systems a larger circuit to assign one or more values to one or more latch variables associated with the larger circuit, generating by the one or more computer systems one or more reduced circuits from the larger circuit according to the values assigned to the latch variables, generating by the one or more computer systems a transition relation (TR) for each reduced circuit, and generating by the one or more computer systems an initial state set for one or more instances of validation on the reduced circuits according to the TRs.
-
System and method for evaluating an erroneous state associated with a target circuit
Issued 7788556
A method for evaluating an erroneous state associated with a target circuit is provided that includes receiving information associated with a target circuit, the information identifying a property within the target circuit to be verified. One or more partitioned ordered binary decision diagram (POBDD) operations are executed using the information in order to identify an erroneous state associated with a sub-space within the target circuit. A path associated with the erroneous state is…
A method for evaluating an erroneous state associated with a target circuit is provided that includes receiving information associated with a target circuit, the information identifying a property within the target circuit to be verified. One or more partitioned ordered binary decision diagram (POBDD) operations are executed using the information in order to identify an erroneous state associated with a sub-space within the target circuit. A path associated with the erroneous state is identified. The path reflects a correlation between an initial state associated with the erroneous state and a point where the erroneous state was encountered.
-
Circuit verification
Issued 7571403
In one embodiment, a method for verifying one or more particular properties of a circuit using a learning strategy to determine suitable values of particular verification parameters includes classifying each of multiple properties of a circuit according to circuit size and selecting a candidate property from the properties. The candidate property set includes one or more particular properties from each property class. The method also includes attempting to verify one or more particular…
In one embodiment, a method for verifying one or more particular properties of a circuit using a learning strategy to determine suitable values of particular verification parameters includes classifying each of multiple properties of a circuit according to circuit size and selecting a candidate property from the properties. The candidate property set includes one or more particular properties from each property class. The method also includes attempting to verify one or more particular properties of the circuit using the candidate property set and particular values of particular verification parameters. The method also includes determining suitable values of the particular verification parameters according the attempted verification of the particular properties of the circuit using the candidate property set and the particular values of the particular verification parameters.
-
Validating one or more circuits using one of more grids
Issued 7546563
In one embodiment, a method for validating one or more circuits using one or more grids includes accessing a circuit and generating one or more seeds for executing one or more instances of validation on the circuit. Each instance of validation comprising one or more tasks. The method also includes distributing the tasks and the seeds across a grid including multiple nodes and, using the seeds, executing the instances of validation at the nodes in the grid according to the tasks.
-
Determining one or more reachable states in a circuit using distributed computing and one or more partitioned data structures
Issued 7216312
In one embodiment, a method for determining one or more reachable states in a circuit using distributed computing and one or more partitioned data structures includes, at a first one of multiple computing systems, receiving a first partition of a circuit. The first partition corresponds to a first binary decision diagram (BDD) having a first density. The method includes performing a first reachability analysis on the first partition using the first BDD until a fixed point in the first partition…
In one embodiment, a method for determining one or more reachable states in a circuit using distributed computing and one or more partitioned data structures includes, at a first one of multiple computing systems, receiving a first partition of a circuit. The first partition corresponds to a first binary decision diagram (BDD) having a first density. The method includes performing a first reachability analysis on the first partition using the first BDD until a fixed point in the first partition has been reached and, if, during the first reachability analysis, the size of the first BDD exceeds a threshold, discarding the first BDD. The method includes communicating with at least one second one of the multiple computing systems. The second one of the multiple computing systems has received a second partition of the circuit. The second one of the multiple computing systems has performed a second reachability analysis on the second BDD without discarding the second BDD.
-
System and method for executing image computation associated with a target circuit
Issued 7032197
A method for verifying a property associated with a target circuit is provided that includes receiving information associated with a target circuit, the information identifying a property within the target circuit to be verified. One or more operations may be executed in order to generate a set of transition relations for performing a reachability analysis associated with the target circuit. An image associated with the target circuit may be partitioned into a plurality of leaves that may each…
A method for verifying a property associated with a target circuit is provided that includes receiving information associated with a target circuit, the information identifying a property within the target circuit to be verified. One or more operations may be executed in order to generate a set of transition relations for performing a reachability analysis associated with the target circuit. An image associated with the target circuit may be partitioned into a plurality of leaves that may each represent a subset of a final image to be generated by a partitioned ordered binary decision diagram (POBDD) data structure. An analysis may be computed of one or more of the leaves using a selected one or both of conjunction and quantification operations separately.
-
Circuit verification
Issued 7028279
In one embodiment, a system for verifying a circuit using a scheduling technique includes one or more partitioned ordered binary decision diagram (POBDD) modules that collectively generate one or more POBDDs. Each POBDD corresponds to one or more partitions of a state space of the circuit and includes a number of states and a number of nodes in the partition. The system also includes one or more cost metrics modules that collectively determine a processing cost of each of the partitions of each…
In one embodiment, a system for verifying a circuit using a scheduling technique includes one or more partitioned ordered binary decision diagram (POBDD) modules that collectively generate one or more POBDDs. Each POBDD corresponds to one or more partitions of a state space of the circuit and includes a number of states and a number of nodes in the partition. The system also includes one or more cost metrics modules that collectively determine a processing cost of each of the partitions of each of the POBDDs. The system also includes one or more scheduling modules that collectively schedule processing of the partitions of the POBDDs for semiformal verification of a circuit. The schedule is based, at least in part, on the determined processing costs of the partitions of the POBDDs.
-
System and method for verifying a plurality of states associated with a target circuit
Issued 6904578
A method for verifying a property associated with a target circuit is provided that includes receiving information associated with a target circuit, the information identifying a property within the target circuit to be verified. One or more partitioned ordered binary decision diagram (POBDD) operations are then executed using the information in order to generate a first set of states at a first depth associated with a sub-space within the target circuit. Bounded model checking may be executed…
A method for verifying a property associated with a target circuit is provided that includes receiving information associated with a target circuit, the information identifying a property within the target circuit to be verified. One or more partitioned ordered binary decision diagram (POBDD) operations are then executed using the information in order to generate a first set of states at a first depth associated with a sub-space within the target circuit. Bounded model checking may be executed using the first set of states in order to generate a second set of states at a second depth associated with the sub-space within the target circuit. The first set of states may be used as a basis for the second set of states such that the second depth is greater than the first depth.
More activity by Subramanian "Subbu"
-
This is the first of three notes "on transitions". This one deals with #ambition and how one thinks about it. At the end of the last month I left…
This is the first of three notes "on transitions". This one deals with #ambition and how one thinks about it. At the end of the last month I left…
Posted by Subramanian "Subbu" Iyer
Other similar profiles
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore MoreOthers named Subramanian "Subbu" Iyer
1 other named Subramanian "Subbu" Iyer is on LinkedIn
See others named Subramanian "Subbu" Iyer