Digital games represent an appropriate test scenario to investigate an agents ability to detect changes in the bahavior of other agents (trying to prevent them from fulfilling their objectives)...Interesting stuff, where a later chat with one of the authors revealed that they had also considered using the same kinds of techniques to detect ''novelty'' in real life problems like financial markets and ECG signals.
... The Markov chain based algorithm M-DBScan is a successful tool conceived to detect novelties in datastream scenarios.
... The main contribution of the present work is to investigate how to improve the use of M-DBScan, as a tool for detecting behavoir changes in the context of Starcraft games (by using a distinct set of featues, and Markov chains, to represent relevant game stages).
... We propose to use a new architecture denoted ''Two-Stream Fully Convolutional Networks''.. ... in order to detect abnormal spatio-temporal events that could present a security risk...I.e. with more and more security cameras, we need automation to detect abnormal events.
Self-learning methods have flourished significantly in recent years.Using reinforcement learning:
One challenge is to learn to survive in a real or simulated world by solving tasks with little prior knowledge about oneself (the agent) the task, and the environment.
In this paper, the state of the art methods of reinforcement learning (in particular, Q-learning), are analyzed regarding applicability to such problems.
The Q-learning algorithm is completed with replay memories and exploration functions...
In a project, a self-learning agent shall be developed that learns from scratch what its sensors and actuators are doing and how to use them to reach a certain goal.The authors had then moved on to doing tests on Andrychowicz's bitflip environment (See: ''Stanford Reinforcement Exercise'' and misc. Reinforcement Problems programmed in Pytorch). Where the goal in the ''bit-flip'' problem is ''to flip the bits in a bit vector of length n in the same way, so that it matches a target bit vector within n tries''.
...
The agent has to predict which action to perform next (in order) to receive a high amount of reward.
...
(Others have found that) Deep Q-learning (DQN) performed well on several Atari 2600 games. DQN outperformed a linear learning function in 43 out of 49 games, and human game testers in 29 out of 49 games (Mnih, 2013 [1]).
Therefore, algorithms from the area of reinforcement learning algorithms will be considered here...
The main objective of this paper was to analyze which combinations of reinforcement learning algorithms, exploration methods and replay memories are most suitable for discrete and continuous state spaces, as well as action spaces. Tests were performed in a simulated discrete ''bit-flip'' and ''continuous pendulum environment''.Where they end up giving us their suggestions for the best way to use these algorithm.
Enabling robots to learn complex tasks through experience allows us to take a big step into the future ... Writing complex algorithms to control these robots is eliminated because they learn to control themselves. In addition, through repetition, they are able to optimize their behavior. Changes in the environment do not affect them, because they can adapt to them automatically.Certainly, a great talk and very straight-forward inspiring work.
Automated planning deals with the problem of finding a sequence of actions leading from a given state to a desired state. The state-of-the-art automated planning techniques exploit informed forward search guided by heuristics (where the heuristic under-estimates a distance from a state to a goal state).Conclusion:
...
Heuristic learning is a relatively new field, which studies how machine learning (ML) can be used to construct a heuristic used in informed forward-search algorithms, such as A* and IDA*.
...The objective is to train a regression-based ML model to be able to estimate goal-distances of states, and then to use this trained model as a heuristic function during search.
This is a technique to automatically construct a strong heuristic for a given planning domainInd the end, their experiments ended up showing that these techniques are just as good as most popular domain-independent heuristics. Which sounded useful indeed.
(Where a trained network can be used as a heuristic on any problem from the domain of interest...).
How a viewer is guided to the focal point of a painting is not only a matter of visual perception, but also of interpretation and meaning making (Painters deliberately create focal points based on sophisticated knowledge of human perception and interpretation).
An angel appears to Christ in the left foreground, holding a chalice in his hand. In the right-hand background Judas and a group of soldiers approach to arrest Christ, marching through a dry landscape without vegetation [2].Compare this to (most) other situations, where people only quickly glance at a picture, and doesn't really see anything...
There has been some work on sentiment analysis of visual images, but this has not at all reached the depth with which humans are able to grasp the emotional content in a scene.3) Even more difficult:
An even higher level of interpretation requires ''understanding''. Who or what is actually depicted, in order to understand the motivations and intentions of those being represented, or the situations in which they find themselves.Indeed ...
Understanding this level requires knowledge about many levels of meaning, because it rests on knowing conventions in society and knowing about historical events, well known figures, and cultural artefacts, like books or films.AI's that should be able to deal with such problems will have to be equipped with knowledge representations, reasoning, and semantic memory (knowledge graphs with vast amounts of facts).
The main challenge that artificial intelligence research is facing nowadays is how to guarantee the development of responsible technology... And, in particular, how to guarantee that autonomy is responsible. The social fears on the actions taken by AI can only be appeased by providing ethical certification and transparency of systems.Indeed, ''Responsible AI'' is quite a catch-phrase these days.
AI will be ''social'', and there will be thousands of AI systems interacting among themselves, and with a multitude of humans; (So, ethical AI systems should be capable of deciding what are the social norms, and abide by them...).
AI needs the trust of citizens to develop. To earn this trust, AI will have to respect ethical standards reflecting our values. Decision-making must be understandable and human-centric [3].Which follows an earlier EU ''communication'', where the focus was to Build Trust in Human-Centric Artificial Intelligence, In the ''communication'', it was also stated, that ''European values'' should be at the heart of creating the right environment of trust (for the successful development and use of AI).
(Still) Sierra noted that ''Responsible Behavior'' is a social convention. It is relative, and different for different cultures.
- Human agency and oversight.
- Technical robustness and safety.
- Privacy and Data Governance.
- Transparency.
- Diversity, non-discrimination and fairness.
- Societal and environmental well-being.
- Accountability.
It was long unanimously held among economists that natural resources that were collectively used by their users would be over-exploited and destroyed in the long-term. Elinor Ostrom disproved this idea by conducting field studies on how people in small, local communities manage shared natural resources, such as pastures, fishing waters and forests. She showed that when natural resources are jointly used by their users, in time, rules are established for how these are to be cared for and used in a way that is both economically and ecologically sustainable [5].The idea here is that we could come up with models, where agents not only work together, but also monitor each other.
Aspect-based sentiment analysis aims at finding expressed opinions towards attributes of products or services (Sentiment analysis, or opinion mining, has experienced a growing interest in recent years due to a growing amount of user-generated content and, until recently, a lack of methods to make sense of this content).The authors paper deals with the domain of physician reviews.
Dr. Meyer knows what he is doing [competence] and is cordial [friendliness] and takes time [time taken] for the patient, his explanations are great [explanation].Still, searching for a few words is not much of model.
Our architecture builds mainly on a bidirectional Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) system for extracting features from sequential text data in both directions, using words before and after the current one.Apparently a good choice, as they outperformed a number of other approaches.
- How old are you?are mapped to about the same thing in a word-embedding. I.e. the idea is to ''put the entire sentence into a vector-space''.
and
- What is your age?
Contextual alignment refers to the process of matching incoming utterances against similar utterance patterns within a scaffold corpus.In order to simulate an open-domain dialog setting, they needed a lot of candidate utterances with a broad spread of conversation topics and dialog styles. To achieve this, they merged data from four different sources:
1. Chit-Chat [6].To explore the types of conversations produced by their system, they then allowed a human to type the opening line to give to their algorithm...
2. Daily Dialog [7].
3. A 33 million word subset of Reddit [8].
4. Ubuntu Dialogue Corpus [9].
(Conversation generated using the scattershot scaffolding algorithm. Human-generated text is preceded by the > symbol.)Indeed, loads of fun. And a great presentation.
> How is your day going so far?
Hi, it's good!
> Good to hear. What are you working on?
I'm figuring out my budget .
> Oh that sounds fun. What are you going to do after you finish?
Well, I plan to get into the accounting program.
> Oh cool! Do you like accounting?
I've thought about going into cardiology, that seems interesting, but also I’m considering anesthesiology.
> Oh wow. Those are both very different from accounting. Are they hard?
He also gave a list of AI research priorities in the coming years (As stated by the community):
- Reduced healthcare cost.
- Universal personalized education.
- Evidence-driven social opportunity.
- Accelerated scientific discovery.
- Unprecedented innovation for businesses.
- (Better) National defense and security.
E.g.Earlier, computers could only go so far, as humans had to type in everything they should work with.
(In the area of) Integrated Intelligence:(In the area of) ''Self-aware'' Learning:
- Science of integrated intelligence.
- Understanding human intelligence.
- Robust and trustworthy learning.
- Integrating symbolic and numeric representations.
The action language golog has been used in a large number of applications (as a high-level control language), from intelligent service robots to soccer robots...It all sounded very convincing. Especially, as the Robot Operating System (ROS), in recent years, has emerged as the de-facto standard (For tools to help you build your robot applications).
For the lower level robot software, the Robot Operating System (ROS) has been around for more than a decade now, and it has developed into the standard middleware for robot applications. ROS provides a large number of packages for standard tasks in robotics like localisation, navigation, and object recognition
...
Here, we describe our approach to marry the golog action language with ROS.
In this study, we propose a method for transferring the shading style used on hair in one drawing to another drawing...
For training, we used the Pix2pix network (Isola 2016). Pix2pix is a cGAN-based image-generation neural network technology.
The GAN architecture is comprised of a generator model for outputting new plausible synthetic images, and a discriminator model that classifies images as real (from the dataset) or fake (generated).And most importantly: It actually works here! Cool!
...
The Pix2pix model is a type of conditional GAN, or cGAN, where the generation of the output image is conditional on an input, in this case, a source image.
...
The Pix2pix GAN has been demonstrated on a range of image-to-image translation tasks such as converting maps to satellite photographs, black and white photographs to color, and sketches of products to product photographs.
Birds learning to sing can be compared to infants babbling.Indeed, there is a lot of (hidden) meaning in music.
...
And, the origins of language as a system of signifiers, might have evolved from the imitation of the sounds of various predators, which functioned as warning signs [16].
Recurrent neural networks (RNNs) can be trained to process sequences of signals. Often with quite impressive results within different areas.Still, the predicted signals (in music) aren't always perfect matches. So, in this work, they propose a progressive learning strategy that can mitigate the mistakes by using domain knowledge.
The Annotated Beethoven Corpus (ABC), which contains harmonic analyses of all sixteen Beethoven string quartets. Comprising Beethoven's middle and late creative phrases; and hence, both the high classical and early Romantic eras.Using this domain background knowledge, they do see some performance improvements:
Our experiments on chord progression reveal that our proposed approach can yield predictive performance improvement without incurring longer training time.Very interesting work, indeed.
When engaging in social interaction, people rely on their ability to reason about other people's mental states, including goals, intentions, and beliefs.In order to get us started with this reasoning about other peoples minds, we were introduced to the ''Luc Steels birthday - puzzle''. Here we would only be able guess the correct date, if we were able to correctly deduce what 2 other people (Kim and Harmen) knew about his birthday.
This theory of mind ability allows them to more easily understand, predict, and even manipulate the behavior of others.
In the coming era of hybrid intelligence, in which teams consist of humans, robots and software agents, it will be beneficial if the computational members of the team can diagnose the theory of mind levels of their human colleagues.Of course!
Palermo (Sicily).
