| abstract | link |

Most of the existing algorithms for zero-shot classification problems typically rely on the attribute-based semantic relations among categories to realize the classification of novel categories without observing any of their instances. However, training the zero-shot classification models still requires attribute labeling for each class (or even instance) in the training dataset, which is also expensive. To this end, in this paper, we bring up a new problem scenario: “Are we able to derive zero-shot learning for novel attribute detectors/classifiers and use them to automatically annotate the dataset for labeling efficiency?”. Basically, given only a small set of detectors that are learned to recognize some manually annotated attributes (i.e., the seen attributes), we aim to synthesize the detectors of novel attributes in a zero-shot learning manner. Our proposed method, Zero-Shot Learning forAttributes (ZSLA), which is the first of its kind to the best of our knowledge, tackles this new research problem by applying the set operations to first decompose the seen attributes into their basic attributes and then recombine these basic attributes into the novel ones. Extensive experiments are conducted to verify the capacity of our synthesized detectors for accurately capturing the semantics of the novel attributes and show their superior performance in terms of detection and localization compared to other baseline approaches. Moreover, with using only 32 seen attributes on the Caltech-UCSD Birds-200-2011 dataset, our proposed method is able to synthesize other 207 novel attributes, while various generalized zero-shot classification algorithms trained upon the dataset re-annotated by our synthesized attribute detectors are able to provide comparable performance with those trained with the manual ground-truth annotations.

| abstract | link | project |

In this paper, we proposed a novel task-consistency learning method that allows training a vacant space detection network (target task) based on the logic consistency with the semantic outcomes from a flow-based motion behavior classifier (source task) in a parking lot. By well designing the reward mechanism upon semantic consistency, we show the possibility to train the target network in a reinforcement learning setting. Compared with conventional supervised detection methods, this work's main contribution is to learn a vacant space detector via semantic consistency rather than supervised labels. The dynamic learning property may make the proposed detector been deployed and updated in different lots easily without heavy human loads. The experiments show that based on the task consistency rewards from the motion behavior classifier, the vacant space detector can be trained successfully.

... Read more

[Goal]
Transfer an existing well-performing HVAC (large air conditioning) control agent to another building with a different sensor deployment (i.e., different types and numbers of sensors) and lack of data.

... Read more

[Goal]
To develop a mechanism for training a vacant parking space detection model in a new parking lot without manually annotating or storing data from another parking lot for transfer learning.
[Challenges]
1. The system should be source domain data-free.
2. The system should be target domain label-free.
[Designs]
In order to meet the requirement, instead of transfer learning, we trained a vacant space detection network (target task) based on the logic consistency with the semantic outcomes from an optical-flow-based car motion detector (source task) in a parking lot via reinforcement learning. The key points are summarized as followed:
1. Source domain data-free: At a start, a car motion detection model is trained by the source data, to predict a slot is either in "car in" state, "car out" state, or "no motion" state, given optical flow as the input. Under this setting, the motion detection model is stored for guiding the target task, but not the source dataset with its labels.
2. Target domain label-free: During training the target task (i.e., vacant space detection model for the target parking lot), we defined the reward function based on the logical relation between car motion and parking slot occupancy. In this way, instead of labeling, the car motion detector can assign the reward for the target task to learn.
3. Since the car motion detection model from the source parking lot might be undesirable for samples from the target parking lot, we design a special warm-up trick to adjust the reward function to deal with the noisy rewards.

... Read more

Goal:
To infer the occupancy based on multiple non-image-based sensor values as the indicator for decision-making in building equipment control.
Challenges:
1. To protect the privacy, only non-image-based sensors are available.
2. To provide continuous indicator, given real time sensor status, for multiple mode control, with extremely rough and sparse label. (i.e., only samples that are considered to be either "full occupancy" or "no people" are annotated.)
Designs:
1.Inspired by the design of the discriminator in WGAN, a neural network is trained, with the gradient penalty for Lipschitz continuity, to project the sensor states into [0,1] space. We force the "full occupancy" samples to be projected as far away as possible from those samples labeled as "no people", while the unlabeled samples should be smoothly distributed between these two labeled groups.
2.Since the labeled samples are rare, we additionally add a first-order gradient constraint loss to guide the prior knowledge about the relationship between environmental factors and occupancy. (For example, The amount of change in the power supply may be positively correlated with the occupancy changes.)

... Read more

Goal:
To accurately stabilize the temperature in a large-scale office at the user’s expected by automatically adjusting the control parameter of HVAC in a building with rare and biased data.
Challenges:
1. Collecting data for smart building is extremely time consuming. To work in the real world, the system is required starting with limited.
2. In many cases, some of the control parameters for equpments in the smart building are never changed. These biased data can not well represent the how actually the control parameters interact with the world.
3. In the real-world scenario, the agent should provide acceptable confort quality while exploring in the early stage.
Designs:
To solve this issue, the algorithm is designed to be a dynamic-learning-based model predictive control system. The key designs are summarized as followed:
1. To solve the problem of rare and biased data, a first-order gradient constraint loss is introduced for guiding the model to learn the correct relationship among environment factors. (For example, the change of AC's temperature setpoint should be positively related to the future temperature changes.)
2. To avoid our agent to perform some extreme setpoints in the early stage, we apply a weighted mixup control strategy with a "safe" constant setpoint defined by user. The mix-up weight is dynamically adjusted depend on the model's qulity (i.e., related to predition error of the latest data sample).