[Bmi] The First Artificial Intelligence Machine Learning (AIML) Contest and other BMI 2016 programs

[Juyang Weng]

Dear *C*olleagues:

Happy New Year!

BMI is pleased to announce the first Artificial Intelligence Machine 
Learning (AIML) Contest in the BMI summer 2016 program.  See below.

*The First Artificial Intelligence Machine Learning (AIML) Contest* 
<http://www.brain-mind-institute.org/program-summer-2016.html>*
**BMI Summer School**and the International Conference on Brain-Mind 
2016* <http://www.brain-mind-institute.org/program-summer-2016.html>

*Important dates*:
Monday March 14, 2016: recommendation of learning engines
Monday April 11, 2016: deadline for advance registration of contest entries
Monday April 11, 2016: deadline for application of BMI course-program 
admission
Monday April 25, 2016: deadline for late registration of contest entries
Monday April 25, 2016: deadline for BMI course registration
May 30 - June 17, 2016: distance learning course for the first three weeks
June 20 - July 8, 2016: distance learning course for the second three weeks
July 11 - July 29, 2016: distance learning course for the third three weeks
Aug. 1 - Aug. 14, 2016: workshops (free for all registered players, 
distance or on site)
Monday, August 15, 2016: Performance run by contest entries due by noon
August 20-21, 2016: ICDL 2016: Contest announcements, sponsor awards, 
and contest presentations (on site and webcast)

*The 1st AIML Contest*

The terms such as artificial intelligence, machine learning, robotics, 
signal processing, control, dynamic systems, data mining, big data, and 
brain projects, often had different emphases but the related disciplines 
are converging.  The Artificial Intelligence Machine Learning (AIML) 
Contest serves as a converging platform for all related disciplines and 
beyond.  It is open to, but not limited to, all researchers, 
practitioners, students and investors.   The main goal of the Contest is 
to promote understanding of natural and artificial intelligence, beyond 
the currently popular pattern classification.   The Contest aims to 
address major learning mechanisms in natural and artificial 
intelligence, including perception, cognition, behavior and motivation 
that occur in cluttered real-world environments.  Attention, 
segmentation, emergence of spatiotemporal representations, and 
incremental scaffolding are part of each life-long learning stream.

The major characteristics of this contest include:
(1) Use inspirations from learning by natural brains, such as grounding, 
emerging, natural inputs, incremental learning, real-time and online, 
attention, motivation, and abstraction from raw sensorimotor data.
(2) General purpose learning engines.   Learning engines will be 
available to participants and open for additional learning engines.  The 
providers of learning engines are free to provide assistants to 
participants, such as courses, tutorials, and workshops.
(3) Training-and-testing sensorimotor streams will be provided to the 
participants.  Each frame of the stream contains a sensory vector and a 
motoric vector.  Training and testing are mixed in the streams, so that 
learning systems can perform scaffolding: early learned simpler skills 
are automatically selected and used for learning later more complex skills.
(4) Major AI challenges will be tested, including vision, audition, 
language understanding, and autonomous thinking.
(5) The Contest is open to investors, charities, governments and 
industrial supporters who like to contribute award funds and provide 
assistance to their learning engines.

Rules:  The entry of each contest is uniquely identified by the name of 
the entry system.   A system is developed by a team consisting of one or 
multiple team members.  A person can participate in one or multiple 
teams.  Although the format of supplied streams is meant for incremental 
learning, at this first year of the contest we allow teams to use either 
framewise incremental or block-incremental learning, but the size of 
block must be reported for contest.  During block-incremental learning, 
the system takes a block of b consecutive frames at a time, update the 
system, and then discard the block.  Framewise incremental learning has 
a block size b=1 frame.  Each system can also run each training stream a 
few times as practice (epochs).  The number of epochs is also reported 
for the Contest.  Entries are submitted via Internet and no travel is a 
must.   The Contest will provide software interface for 
training-and-testing.   Organizers of the contest are ineligible for 
team members of any entry.

The International Conference on Brain-Mind (ICBM) 2016 will feature 
Contest score announcement, sponsor rewards, and team presentations.

Criteria of performance in the following priority of importance (1 is 
the highest):
(1) average error rates over all test points during epoch e, e = 1, 2, ...
(2) the block size is as small as possible to reach a state-of-art error 
rate.
(3) the number of practice is as small as possible to reach a 
state-of-art error rate.
(4) the size of the network is as small as possible to reach a 
state-of-art error rate.

Within each stream, the following five types of substreams (each 
contains multiple tasks and subtasks, skills and subskills) will be 
trained and tested on but each team is not told which type a substream 
is.  It is a violation of the contest rules to manually browse through 
the stream to find out what type a stream is.  The Contest software will 
record all the training and testing data.

Type 1: Spatially non-attentive and non-temporal streams: many 
components of a sensory frame are related to the next motoric frame 
(e.g., the object of interest almost fills the entire image and the next 
motoric frame contains the object type). Non-temporal here means that a 
single frame is sufficient to decide the next motor frame.   This is 
similar to monolithic pattern classification (e.g., image 
classification). But past experience is useful for later learning within 
the same training-and-testing stream.

Type 2: Spatially attentive and non-temporal streams: a relatively small 
number of components of a sensory frame are related to the next motoric 
frames (e.g., the car to be recognized and detected is in a large 
cluttered street scene where the next motoric frames should contain the 
location, type, and scale of the attended car).  Type 2 is a spatial 
generalization of Type 1.  This is like object recognition and detection 
from cluttered dynamic scenes conducted concurrently (where the next 
motoric frames provide desired actions).   Each sensory frame is not 
segmented but internal automatic segmentation needs to be learned.  
Namely, skills to find which image patch is related to the action in the 
motoric frame need to be gradually learned from earlier learning and 
refined in later learning within the same stream.   The early attention 
skills can be learned from motor vector (supervised learning) and/or 
through reinforcement learning (pain and sweet signals in sensory 
frames).  The motoric frames may contain action-supervision signals and 
the sensory frames may contain components for reinforcement signals 
(rewards or punishment components like pain receptors and sweet 
receptors).   The contents in each sensorimotor frame signal what 
learning modes are needed.   For example, a supplied action in a motoric 
vector calls for supervised learning, a supplied pain signal in a 
sensory vector calls for reinforcement learning, and the presence of 
both calls for a combination of supervised learning and reinforcement 
learning.

Type 3: Spatially non-attentive and temporal steams: each motoric frame 
is a function of not only the last sensory frame but also an unknown 
number of earlier sensory frames.
Each motoric frame corresponds to the temporal state/action.  Type 3 is 
a temporal generalization of Type 1.  This is like recognizing sentences 
from a TV screen where the TV screen presents one letter at a time.  
Again, past experience is useful for later learning (e.g., learning 
individual letters and punctuations, individual words, individual 
phrases, individual sentences, etc. progressively, through a single long 
stream).

Type 4: Spatially attentive and temporal steams: each motoric frame is 
related to parts of recent sensory frames.  Type 4 is the temporal 
generalization of Types 2 and the spatial generalization of Type 3.  An 
example is recognizing and detecting the intent of a car moving in a 
cluttered scene.   Again, earlier experience is useful for later 
learning (e.g., motion direction, motion patterns, object type, object 
location, object orientation, etc.).

Type 5: Generalization that requires certain amount of autonomous 
thinking: the actions in the motoric frame require the system to invent 
rules and use such rules on the fly within the same (long) 
training-and-testing stream. Type 5 is the thinking generalization of 
Type 4.  Classical conditioning, instrumental conditioning, autonomous 
reasoning, and autonomous planning are examples.

Practice streams for training-and-testing will be provided by the 
Contest early on.   For the Contest, each entry is required to run 
through a Contest Interface, which records the performance in real 
time.  The frame rate is around 10Hz in real time, but each entry can 
run slower in virtual time.  GPU is recommended but not required.   The 
information about the computer architecture should be provided.   
Spatial and temporal computational complexities are considered in 
Criteria (3) and (4).

Open-Source Machine Learning Engines available:
(1) Google TensorFlow
(2) MSU Developmental Network (DN)
(3) Submission or recommendation of learning engines for contest: open 
till Monday March 14, 2016
Each supplier or recommender of an engine is free to decide courses and 
workshops below, but such assistance is recommended but not required.

Entries for contest:
Advance registration deadline: Monday April 11, 2016.
Registration: $270 per entry.   Scores are measured based on entries.
Each team can register for multiple entries; a team can register for 
multiple human participants; each participant can register for multiple 
teams.
The first name of each entry is waived of the $90 registration fee for 
three courses/tutorials.
Full-time student players: waived of tuition for courses/tutorials other 
than the $90 registration fee.
Every player to be officially recognized needs to register in the 
Contest Registration Form.
Course registration deadline: Monday April 25, 2016.

Contest subject areas:  Each entry chooses at least one of the following 
four subjects:
(1) vision,
(2) audition (including speech and music recognition),
(3) natural language understanding,
(4) creative machine thinking for one of the above three areas or more.
Each entry can use one or more of the machine learning engines provided 
by the Contest or elsewhere.
Each entry can address one or more challenge areas.

Each supplier of the Machine Learning Engines is encouraged to provide 
courses or tutorials, via BMI or independently.
BMI Courses or Tutorials for Machine Learning Contest engines:
Application for BMI Admission: deadline: Monday April 11, 2016.
Full time students: waive of tuition for courses
Course registration deadline: Monday April 25, 2016.
May 30 - June 17, 2016 (distance learning course for three weeks, 
including BMI 831 <http://www.brain-mind-institute.org/bmi-831.html>)
June 20 - July 8, 2016 (distance learning course for three weeks, 
including BMI 861 <http://www.brain-mind-institute.org/bmi-861.html>)
July 11 - July 29, 2016 (distance learning course for three weeks, 
including BMI 871 <http://www.brain-mind-institute.org/bmi-871.html>)
Aug. 1 - Aug. 14, 2016 workshops (free for all registered players).

Contest entries due: noon, Monday, August 15, 2016.
5-day evaluation.
Contest award meeting and Contest presentations (ICDL 2016): August 
20-21, 2016.

Award amount: $50,000, to be updated with sponsors.

Data:  The training-and-testing streams will be provided.  Many machine 
learning techniques are for off-line, batch training, batch testing, and 
task specific.  They must be modified to take the official 
training-and-testing streams for online training and testing.  Each 
stream consists of a single sequence of many time frames; each time 
frame i contains a sensory frame X[i] and a motoric frame Z[i].   Each 
motoric frame many include both training data points and testing data 
points.   If a motoric frame that is  marked * (free), it is a testing 
frame, absent of training data.   Namely, each stream is a synchronized 
sensorimotor sequenced (X[i], Z[i]), i = 0, 1, 2, … n, where X[i] and 
Z[i] are the sensory vector (e.g., image) and action vector (state) at 
time i, both non-symbolic (numeric vector) to promote fully automatic 
machine learning. Z[i] includes binary components that represent 
abstract concepts of a spatiotemporal event (e.g., location concept, 
type concept, state concept of a sentence). X[i] may include specified 
components as punishments and rewards for action Z[i-1] or a few frames 
earlier (not too much delay that confuses with earlier actions).  There 
are two types of Z[i]’s, supervised and free, respectively.  Namely, 
free Z[i]’s are motor vectors for test.  Each Z[i] consists of a number 
of concept zones [e.g., Z=(ZT, ZL,ZS), where ZT, ZL, ZS represent type 
zone, location zone, and scale zone, respectively for the attended 
object].  With each zone, only one neuron can fire at 1 and all other 
neurons do not fire and take value 0. Within each stream, past learned 
skills with early i’s is useful for later learning with later i’s.

Contest: each entry runs a contest software provided by the BMI for 
training-and-testing.  The performance is recorded and reported by the 
contest software automatically.

*BMI Summer School and ICBM 2006*

2006 is the 5th year of BMI summer school. It is also the first time 
that the summer school is jointly run with the AIML Contest as past of 
the educational support of the Contest.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cse.msu.edu/pipermail/bmi/attachments/20151231/244088dc/attachment-0001.html>